ADVANCES IN ADAPTIVE CONTROL THEORY:

GRADIENT– AND DERIVATIVE–FREE APPROACHES

A Dissertation Presented to
The Academic Faculty of
The School of Aerospace Engineering

by

Tansel Yucelen

In Partial Fulfillment of
The Requirements for the Degree of
Doctor of Philosophy in Aerospace Engineering

Georgia Institute of Technology

May 2012
 ADVANCES IN ADAPTIVE CONTROL THEORY:
GRADIENT– AND DERIVATIVE–FREE APPROACHES

Approved by:

Professor Anthony J. Calise Eugene Lavretsky, Ph.D.

Committee Chair and Advisor Senior Technical Fellow
School of Aerospace Engineering The Boeing Company
Georgia Institute of Technology

Professor Eric Feron Professor Panagiotis Tsiotras

School of Aerospace Engineering School of Aerospace Engineering
Georgia Institute of Technology Georgia Institute of Technology

Professor Eric N. Johnson Date Approved: 14 April 2011

School of Aerospace Engineering
Georgia Institute of Technology
To my wife Gulfem Ipek Yucelen,

the origin of my love and passion.

iii
 ACKNOWLEDGEMENTS

It is my great pleasure to take this opportunity to thank the people who directly and
indirectly played a key role in successful completion of this work. I would like to thank

first to my advisor Dr. Anthony J. Calise for his support, advice, and guidance over

the years that I have been working with him. His wisdom and insight has helped me

grow as a researcher and engineer. While I was doubtful and my idea was superficial,

he could concretely envision what would be achieved, and, without him, I could have

never come to this stage. In addition, his ability to find innovative solutions above
and beyond the convention will always inspire me.

I would also like to thank Dr. Wassim M. Haddad for the many interesting

discussions on control theory. I greatly enjoyed our conversations on different subjects.

He inspired and broadened my view of control research from the mathematical aspect.

He has made a deep impact on the rigor of my research. Furthermore, I always feel

indebted for his excellent teaching.

Dr. Eric N. Johnson has taught me to appreciate the value of intuition and insight

in control theory. Our discussions with him helped me to understand and envision

my knowledge of applied control theory. Therefore, I thank to him and his graduate

research assistants in the unmanned aerial vehicle research facility, who helped me to

implement the theories of this dissertation to real world systems.

I want to thank many people at NASA that I worked with on the Generic Trans-
port Model and AirSTAR including Dr. Nhan T. Nguyen, Dr. David Cox, Dr. Kevin

Cunningham, and Dr. Austin M. Murch. I learned a lot from our interactions.

Moreover, without their help, AirSTAR applications would have been missing in this

dissertation.

iv
 I am also indebted to Dr. Eugene Lavretsky, Dr. Panagiotis Tsiotras, and Dr.
Eric Feron. Their valuable suggestions and insights greatly improved the quality of

the thesis. I learned a lot from the review process.

I am thankful for all of the enlightening discussion with all of my co-workers at

Georgia Tech. I would like to thank Kilsoo Kim, Rajeev Chandramohan, Jonathan

Muse, Girish Chowdhary, Ali Kutay, Konstantin Volyanskyy, and Suresh Kannan.

In addition, I must say that I miss many discussions that I had with Dr. Bong-Jun
Yang. He has been a tremendous help in my earlier studies at Georgia Tech.

I am also grateful to Dr. Abdulrahman H. Bajodah from King Abdulaziz Univer-

sity, Dr. Anuradha Annaswamny from Massachusetts Institute of Technology, Arjun

Sadahalli from Southern Illinois University Carbondale, Dr. Christine M. Belcastro

from NASA, Dr. Farshad Farid from Whirlpool Corporation, Dr. Farzad Pour-

boghrat from Southern Illinois University Carbondale, Dr. Florian Holzapfel from
Technical University of Munich, Dr. Gang Tao from University of Virginia, Dr. Irene

M. Gregory from NASA, Dr. John J. Burken from NASA, Dr. Jong-Yeob Shin from

Gulfstream Aerospace Corporation, Dr. Kevin Wise from Boeing, Dr. Luis Crespo

from NASA, Dr. Margareta Stefanovic from University of Wyoming, Dr. Mark Balas

from University of Wyoming, Dr. Naira Hovakimyan from University of Illinois at

Urbana Champaign, Dr. Peda V. Medagam from Phase Technologies LLC, Dr. S.
N. Balakrishnan from Missouri University of Science and Technology, Dr. Tannen

Vanzwithen from NASA, Dr. Quang M. Lam from Orbital Sciences Corporation, Dr.

Vahram Stepanyan from NASA, and Dr. Warren Dixon from University of Florida,

for many interesting discussions and/or suggestions.

Finally, I would like to express my deepest gratitude to my wife Ipek for her never

ending love and support. Without her, I would not have achieved this work like many
of the other things that I have in my life.

v
 TABLE OF CONTENTS

Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv

I Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Model Reference Adaptive Control . . . . . . . . . . . . . . . . . . . 1
1.2 Linear Constraints and Modification Terms . . . . . . . . . . . . . . 2
1.3 Linear Constraints and Weight Update Laws . . . . . . . . . . . . . 4
1.4 Derivative-Free Adaptive Control . . . . . . . . . . . . . . . . . . . . 4
1.5 Output Feedback Adaptive Control . . . . . . . . . . . . . . . . . . 7
1.6 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

II Kalman Filter Modification in Adaptive Control . . . . . . . . . . 10

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Approximate Enforcement of a Linear Constraint by KF Optimization 17
2.4 Examples on a Model of Wing Rock Dynamics . . . . . . . . . . . . 20
2.4.1 A KF Alternative to e–Modification . . . . . . . . . . . . . . 21
2.4.2 A KF Alternative to ALR–Modification . . . . . . . . . . . . 30
2.4.3 A New ue –Modification for Input Constraints . . . . . . . . . 35
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

III Kalman Filter Based Adaptive Control . . . . . . . . . . . . . . . . 41

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

vi
 3.3 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Kalman Filter Based Adaptive Control Formulation . . . . . . . . . 45
3.5 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6 Examples on a Model of Wing Rock Dynamics . . . . . . . . . . . . 49
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

IV Derivative-Free Adaptive Control: The Full-State Feedback Case 55

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3 Derivative-Free Adaptive Control . . . . . . . . . . . . . . . . . . . . 56
4.4 Modifications to Derivative-Free Adaptive Control . . . . . . . . . . 66
4.5 Extensions to the Input Uncertainty Case . . . . . . . . . . . . . . . 69
4.6 Examples on a First Order System . . . . . . . . . . . . . . . . . . . 73
4.6.1 Uncertainty with Time-Varying Ideal Weights . . . . . . . . . 73
4.6.2 Uncertainty with Constant Ideal Weights . . . . . . . . . . . 79
4.6.3 Input Uncertainty Case . . . . . . . . . . . . . . . . . . . . . 81
4.7 Examples on a Generic Transport Model . . . . . . . . . . . . . . . 83
4.7.1 Missing Left Wing Tip . . . . . . . . . . . . . . . . . . . . . 85
4.7.2 Missing Vertical Tail . . . . . . . . . . . . . . . . . . . . . . 89
4.8 Flight Test Results on NASA AirSTAR . . . . . . . . . . . . . . . . 92
4.8.1 Case 1: Employing Baseline and Adaptive Controllers . . . . 92
4.8.2 Case 2: Roll and Pitch Rate Commands . . . . . . . . . . . . 92
4.8.3 Case 3: Latency Emulation . . . . . . . . . . . . . . . . . . . 93
4.8.4 Case 4: Cmα and Clp Reduction by 50% and +0.2 under Pitch
Rate Command . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.8.5 Case 5: Cmα and Clp Reduction by 50% and +0.2 under Roll
Rate Command . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.8.6 Case 6: Cmα and Clp Reduction by 75% and +0.3 under Pitch
Rate Command . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

vii
V Derivative-Free Adaptive Control: The Output Feedback Case 99
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2 Derivative-Free Output Feedback Adaptive Control Architecture . . 100
5.3 A Parameter Dependent Riccati Equation . . . . . . . . . . . . . . . 103
5.4 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.5 Examples on a Model of Wing Rock Dynamics . . . . . . . . . . . . 113
5.6 Examples on an Aeroelastic Generic Transport Model . . . . . . . . 114
5.6.1 Nonlinear Uncertainty . . . . . . . . . . . . . . . . . . . . . . 126
5.6.2 External Disturbance . . . . . . . . . . . . . . . . . . . . . . 128
5.6.3 Nonlinear Uncertainty and External Disturbance . . . . . . . 130
5.6.4 Sudden Change in Ap . . . . . . . . . . . . . . . . . . . . . . 131
5.6.5 Sudden Change in Cmα . . . . . . . . . . . . . . . . . . . . . 133
5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

VI Concluding Remarks and Future Research . . . . . . . . . . . . . 138

6.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.2 Recommended Future Research . . . . . . . . . . . . . . . . . . . . . 139

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

Vita . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

viii
 LIST OF TABLES

1 DF-MRAC laws for various modification terms . . . . . . . . . . . . . 68

ix
 LIST OF FIGURES

1 Augmenting adaptive control of a baseline control system. . . . . . . 13

2 Geometric representation of sets. . . . . . . . . . . . . . . . . . . . . 15
3 Baseline control response without and with uncertainty. . . . . . . . . 25
4 Comparison of adaptive controller responses without modification, with
e-modification, and with KF based e- modification with all gains set
to 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5 Comparison of adaptive controller responses without modification, with
e-modification, and with KF based e- modification with all gains set
to 25. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6 Comparison of adaptive controller responses without modification, with
e-modification, and with KF based e- modification with all gains set
to 100. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
7 Comparison of adaptive controller responses without modification, with
e-modification, and with KF based e- modification with all gains set
to 250. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
8 Comparison of du(t)/dt responses without modification, with e-modification,
and with KF based e- modification with all gains set to 250. . . . . . 28
9 Comparison of adaptive controller responses under sensor noise without
modification, with e-modification, and with KF based e- modification
with all gains set to 25. . . . . . . . . . . . . . . . . . . . . . . . . . . 28
10 Comparison of du(t)/dt responses under sensor noise without modifi-
cation, with e-modification, and with KF based e- modification with
all gains set to 25. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
11 Standard ALR based adaptive control response with uncertainty using
Eq. (33) with γ = 10 and kgrad = 10. . . . . . . . . . . . . . . . . . . 31
12 Standard ALR based adaptive control response with uncertainty using
Eq. (33) with γ = 100 and kgrad = 10. . . . . . . . . . . . . . . . . . . 32
13 Standard ALR based adaptive control response with uncertainty using
Eq. (33) with γ = 100 and kgrad = 100. . . . . . . . . . . . . . . . . . 32
14 KF-ALR based adaptive control response with uncertainty using Eq.
(34) with γ = 10 and kgrad = 10. . . . . . . . . . . . . . . . . . . . . . 33
15 KF-ALR based adaptive control response with uncertainty using Eq.
(34) with γ = 100 and kgrad = 10. . . . . . . . . . . . . . . . . . . . . 33

x
16 KF-ALR based adaptive control response with uncertainty using Eq.
(34) with γ = 100 and kgrad = 100. . . . . . . . . . . . . . . . . . . . 34
17 Adaptive control response without and with actuator dynamics. . . . 37
18 Performance of adaptive controller with hedging, with actuator dynamics. 37
19 Response of reference model without and with hedging, with actuator
dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
20 Performance of adaptive controller with ue -modification, with actuator
dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
21 Comparison of adaptive control responses with hedging versus with
ue -modification, with actuator dynamics and limits. . . . . . . . . . . 39
22 Performance of adaptive controller with ue -modification, with actuator
dynamics and time delay. . . . . . . . . . . . . . . . . . . . . . . . . . 39
23 TwinSTAR flight test vehicle. . . . . . . . . . . . . . . . . . . . . . . 40
24 Nominal control response. . . . . . . . . . . . . . . . . . . . . . . . . 50
25 Standard e- modification based adaptive controller, γ = 1 and γ ×σ = 40. 51
26 Standard e- modification based adaptive controller, γ = 10 and γ ×σ =
40. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
27 Standard e- modification based adaptive controller, γ = 100 and γ ×
σ = 40. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
28 Standard e- modification based adaptive controller, γ = 1000 and γ ×
σ = 40. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
29 KFe –AC law, σ = 40. . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
30 Adaptation gain Γ(t) for the KFe –AC law in Figure 29. . . . . . . . 53
31 Visualization of the proposed DF-MRAC architecture. . . . . . . . . 60
32 r/δ ∗ versus cr for κ1 = {0.01, 0.02, 0.03} and κ2 ∈ [1, 10]. . . . . . . . 63
33 r/δ ∗ versus cr for κ1 = {0.1, 0.2, 0.3} and κ2 ∈ [1, 10]. . . . . . . . . . 64
34 Responses with nominal controller for the square wave ideal weight. . 75
35 Responses with standard MRAC using γ = 102 and γ = 104 . . . . . . 75
36 Responses with DF-MRAC for the square wave ideal weight. . . . . . 76
37 Responses with standard MRAC using γ = 102 and γ = 104 . . . . . . 76
38 Responses with DF-MRAC for a sinusoidal ideal weight. . . . . . . . 77
39 Expanded view of the estimate of the ideal weight. . . . . . . . . . . 77

xi
40 Responses with DF-MRAC for a band limited white noise ideal weight. 78
41 Expanded view of the estimate of the ideal weight. . . . . . . . . . . 78
42 Responses with standard MRAC for an uncertainty with constant ideal
weights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
43 Responses with DF-MRAC for an uncertainty with constant ideal weights. 80
44 Responses with DF-MRAC for input uncertainty case when (76) is
employed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
45 Responses with DF-MRAC for input uncertainty case when (76), (125),
and (126) are employed. . . . . . . . . . . . . . . . . . . . . . . . . . 82
46 GTM nominal control response for nominal operating conditions. . . 86
47 GTM nominal control response for the missing left wing tip case. . . 87
48 GTM DF-MRAC response for the missing left wing tip case. . . . . . 87
49 GTM DF-MRAC response for the missing left wing tip case with 0.01
seconds of time delay in the rudder channel. . . . . . . . . . . . . . . 88
50 GTM DF-MRAC response with ALR modification term for the missing
left wing tip case with 0.01 seconds of time delay in the rudder channel. 88
51 GTM nominal control response for the missing vertical tail case. . . . 89
52 GTM DF-MRAC response for the missing vertical tail case. . . . . . 90
53 GTM DF-MRAC response for the missing vertical tail case with 0.07
seconds of time delay in the right aileron channel. . . . . . . . . . . . 90
54 GTM DF-MRAC response with ALR modification term for the missing
vertical tail case with 0.07 seconds of time delay in the right aileron
channel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
55 Flight test vehicle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
56 Comparison of baseline and adaptive controller responses for Case 1. . 95
57 Comparison of baseline and adaptive controller responses for Case 2. . 95
58 Comparison of baseline and adaptive controller responses for Case 3. . 96
59 Comparison of baseline and adaptive controller responses for Case 4. . 96
60 Comparison of baseline and adaptive controller responses for Case 5. . 97
61 Comparison of baseline and adaptive controller responses for Case 6. . 97
62 Derivative-free output feedback adaptive control architecture . . . . . 103

xii
63 Nominal and adaptive control responses for the case of constant ideal
weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
64 Nominal and adaptive control responses for the case of time varying
ideal weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
65 Depiction of d(t) and w(t) . . . . . . . . . . . . . . . . . . . . . . . . 116
66 Nominal and adaptive control responses with disturbances for the case
of time varying ideal weights . . . . . . . . . . . . . . . . . . . . . . . 116
67 ∆(t, x(t)) and uad (t) . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
68 Derivative-free output feedback adaptive control architecture . . . . . 120
69 Frequency response of the linearized model . . . . . . . . . . . . . . . 123
70 Frequency response of the loop transfer functions with the loop broken
at the plant input for both full-state feedback and LQG/LTR loops . 124
71 Pitch rate and elevator responses with nominal control in the absence
of uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
72 Measurement and state responses and their estimates with nominal
control in the absence of uncertainty . . . . . . . . . . . . . . . . . . 125
73 Pitch rate and elevator responses with nominal control for the case of
nonlinear uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
74 Pitch rate and elevator responses with adaptive control for the case of
nonlinear uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
75 Pitch rate and elevator responses with nominal control for the case of
external disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
76 Pitch rate and elevator responses with adaptive control for the case of
external disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
77 Pitch rate and elevator responses with adaptive control for the case of
nonlinear uncertainty and external disturbance . . . . . . . . . . . . . 130
78 Pitch rate and elevator responses with nominal control for the case of
sudden change in the sign of second row of Ap . . . . . . . . . . . . . 131
79 Pitch rate and elevator responses with adaptive control for the case of
sudden change in the sign of second row of Ap . . . . . . . . . . . . . 132
80 Measurement and state responses and their estimates with adaptive
control for the case of sudden change in the sign of second row of Ap 132
81 Pitch rate and elevator responses with nominal control for the case of
sudden change in Cmα . . . . . . . . . . . . . . . . . . . . . . . . . . 133

xiii
82 Pitch rate and elevator responses with adaptive control for the case of
sudden change in Cmα . . . . . . . . . . . . . . . . . . . . . . . . . . 134
83 Measurement and state responses and their estimates with adaptive
control for the case of sudden change in Cmα . . . . . . . . . . . . . . 134
84 Pitch rate and elevator responses with nominal control for the case of
sudden change in the sign of Cmα . . . . . . . . . . . . . . . . . . . . 135
85 Pitch rate and elevator responses with adaptive control for the case of
sudden change in the sign of Cmα . . . . . . . . . . . . . . . . . . . . 135
86 Measurement and state responses and their estimates with adaptive
control for the case of sudden change in the sign of Cmα . . . . . . . . 136

xiv
 SUMMARY

In this dissertation, we present new approaches to improve standard designs

in adaptive control theory, and novel adaptive control architectures.

We first present a novel Kalman filter based approach for approximately enforcing

a linear constraint in standard adaptive control design. One application is that this

leads to alternative forms for well known modification terms such as e–modification.

In addition, it leads to smaller tracking errors without incurring significant oscillations

in the system response and without requiring high modification gain. We derive
alternative forms of e– and adaptive loop recovery (ALR–) modifications.

Next, we show how to use Kalman filter optimization to derive a novel adaptation

law. This results in an optimization-based time-varying adaptation gain that reduces

the need for adaptation gain tuning.

A second major contribution of this dissertation is the development of a novel

derivative-free, delayed weight update law for adaptive control. The assumption of
constant unknown ideal weights is relaxed to the existence of time-varying weights,

such that fast and possibly discontinuous variation in weights are allowed. This

approach is particulary advantageous for applications to systems that can undergo a

sudden change in dynamics, such as might be due to reconfiguration, deployment of

a payload, docking, or structural damage, and for rejection of external disturbance

processes.
As a third and final contribution, we develop a novel approach for extending all the

methods developed in this dissertation to the case of output feedback. The approach

is developed only for the case of derivative-free adaptive control, and the extension

of the other approaches developed previously for the state feedback case to output

xv
feedback is left as a future research topic.
The proposed approaches of this dissertation are illustrated in both simulation

and flight test.

xvi
 CHAPTER I

Introduction

1.1 Model Reference Adaptive Control

Research in adaptive control is motivated by the fact that models employed in control

system design may not adequately represent the actual system dynamics due to ide-
alized assumptions, linearization, model order reduction, external disturbances, and

degraded modes of operation. Although robust control design approaches can deal

with these sources of uncertainty, they may fail to satisfy performance requirements

in the face of high levels of uncertainty. On the other hand, adaptive controllers are

able to deal with uncertainty without necessarily sacrificing performance. Further-

more, they require less modeling information than do robust controllers. These facts
make adaptive control theory important for engineering applications.

Adaptive controllers can be classified as either direct or indirect. Direct adaptive

controllers adapt feedback gains in response to system variations without requiring

a parameter estimation algorithm. This property distinguishes them from indirect

adaptive controllers which employ an estimation algorithm to estimate the unknown

system parameters and adapt the controller gains. This dissertation presents develop-
ments for the well known and the most important class of direct adaptive controllers,

namely model reference adaptive controllers (MRAC) [4, 29, 33, 36, 38, 44, 49, 61, 68,

71–73].

MRAC has three major components: Reference model, weight (gain) update law,

and controller. The reference model specifies the desired behavior of the closed-loop

system. The output (resp., state) of an uncertain system is compared to the output

1
(resp., state) of the reference model. This comparison results in an error signal used
in the weight update law. The controller employs the weight information from the

weight update law to form the adaptive control signal.

MRAC was developed by Whitaker and his colleagues around 1958 [4]. Ref. 64

presented this novel idea based on a gradient method. In this reference, the weight

update law is constructed as the negative gradient of a cost function chosen as the

square of the norm of the error signal. This procedure drives the trajectories of the
uncertain system to the trajectories of the reference model as time goes to infinity.

Butchart and Shackcloth [9] and Parks [65] analyzed the stability of this gradient

method for the first time using Lyapunov stability theory [54, 55]. Other notable

early achievements on MRAC can be found in Refs. 25,26,28,57,58,62 and references

therein.

In this dissertation, we present not only new approaches to improve standard

MRAC designs but also novel adaptation laws and MRAC architectures.

1.2 Linear Constraints and Modification Terms

In the literature, adaptation laws that impose constraints on the weights to improve

an existing adaptive law are commonly referred as composite adaptation [71]. In

general, the modification terms are found by taking the gradient of a norm of the

constraint violation. However, using a gradient method can result in slow parameter

convergence towards a local minimum [8]. In addition, modification terms that are

gradient based have a fixed adaptation gain, that often have to be chosen large to
obtain satisfactory results. The use of a high modification gain can interact negatively

with unmodeled dynamics, and amplify the effect of sensor noise.

Many modification terms are reported in the literature [10, 20, 37, 47, 60, 63, 75–

77, 84, 90, 91]. Included among these, σ–modification [37] adds pure damping to

2
the adaptive weight update law and turns it into a lag filter. The negative aspect
about σ–modification term is that it inhibits the adaptation process. By contrast,

e–modification adds variable damping that is proportional to the norm of the error

signal. When the error signal is large, it also inhibits the adaptation process. The

positive aspect of both σ– and e–modification terms is that they ensure a bounded

weight history.

Concurrent learning [20] uses current and past data concurrently in the adaptation
process. It allows the adaptation law to continually train in the background based

on past data while still being responsive to dynamic changes based on the current

data. In this way, concurrent learning incorporates long term learning. Q modification

[75–77,90] is similar in spirit to concurrent learning in its intent to improve adaptation

performance by using a moving window of the integrated system uncertainty. There

is an optimal control theory based modification that improves adaptation in the

presence of large adaptive gain [63]. An adaptive loop recovery (ALR) approach

[10] has been introduced as a modification with the objective of recovering the loop

transfer properties of the reference model. More recently, K modification [47] has

been proposed to add stiffness to the weight update law. When it is used with σ– or

e–modification terms, the transient response of the weight history can be controlled

by selection of a natural frequency and damping ratio.

In this dissertation all of these modification terms are viewed as having been in-

troduced to reduce the violation of a linear constraint on the weights in an adaptive

control algorithm. With this perspective in mind, a Kalman filter (KF) optimiza-

tion method is developed to arrive at alternative forms for all previously mentioned

methods of modification. We develop this approach in the context of approximately

enforcing a linear constraint in a standard MRAC design. The approach is shown to

result in smaller tracking errors without requiring a high modification gain. The ap-

proach is illustrated using a simplified model of aircraft roll dynamics at high angles

3
of attack.

1.3 Linear Constraints and Weight Update Laws

In the previous section we stated that the KF approach can be used for approximately

enforcing a linear constraint in a standard MRAC design which leads to alternative

forms for well known modification terms. An extension of this approach results in a

novel Kalman filter based weight update law for adaptive control. We first show that

the standard MRAC weight update law can be viewed as enforcing a linear constraint

using gradient optimization. Then, we use KF optimization to replace the standard

weight update law with its KF form. This approach results in an optimal time-varying

adaptation gain that allows one to achieve a given performance criteria without the

need for excessive adaptation gain tuning. The simulation results show that it leads

to smaller tracking errors without incurring high frequency oscillations in the system

response.

1.4 Derivative-Free Adaptive Control

Standard MRAC approaches are based on Lyapunov stability theory and either as-

sume or derive a weight update law in the form of an ordinary differential equation for

the weight estimates. All these methods have in common the underlying assumption

that there exists a constant, but unknown, ideal set of weights. Although this assump-

tion seems reasonable and these MRAC architectures work well on many systems, in

some failure modes they may require the use of unrealistically high adaptation gain,
or may fail to achieve the desired level of performance in terms of failure recovery.

MRAC laws that require high gain can excite unmodeled dynamics, typically exhibit

an excessive amount of control activity [81,82], amplify the effect of sensor noise, and

are not sufficiently robust to time delay [63].

4
 In this dissertation, we develop a derivative-free model reference adaptive control
(DF-MRAC) law, which uses the information of delayed weight estimates and the

information of current system states and errors [16, 86–89]. The proposed method is

an extension of the iterative learning approach adopted in Ref. 17 for purposes of

adaptive observer design. We relax the assumption of constant unknown ideal weights

to the existence of time-varying weights, such that fast and possibly discontinuous

variation in weights are allowed. The DF-MRAC law is advantageous for applica-
tions to systems that can undergo a sudden change in dynamics. We prove that the

error dynamics are uniformly ultimately bounded using a Lyapunov-Krasovskii func-

tional [31], without employing modification terms in the adaptive law. We consider

constant unknown ideal weights as a special case and show that the state tracking

error dynamics are asymptotically stable. Finally, we discuss employing various modi-

fication terms for further improving the performance and robustness of the adaptively
controlled system.

DF-MRAC differs from MRAC in that it does not make use of integration in

its weight update law. In fact, DF-MRAC challenges the conventional wisdom of

expecting an adaptive law to cancel the effect of matched constant disturbances and

uncertainties, since one does not need adaptation to correct for biases. Biases can

more reliably be handled using a non-adaptive controller with integral action. Most
conventional flight control systems employ integral action for this purpose. Therefore

it is advantageous when augmenting an existing flight control system with an adaptive

element, that the adaptive law not employ an integral function of the tracking error.

Otherwise, there is the potential for a conflict to arise between the existing controller

and the augmenting adaptive element. The nature of this conflict is often seen as a

slowly varying tracking error that is difficult to remove by adjusting the parameters
in the adaptive law. On the other hand, if for some reason it is desirable to treat

the effect of biases with a standard MRAC weight update law, then in this case the

5
underlying existing (nominal) controller being augmented with an adaptive element
should employ only proportional control. However, it is highly advantageous that

the nominal part of the control contain integral action since this ensures that the

errors in the regulated output variables go to zero for constant disturbances and

constant model errors even if they are unmatched, so long as the closed loop system

remains stable. Therefore, since many existing guidance and flight controllers do

employ integral action, and if the goal is to augment an existing controller, then it is
not desirable that the weight update law have the effect of integral action. In these

circumstances, the role of adaptation should be confined to things like maintaining

stability and error transient performance, fast upset recover in the event there is a

destabilizing failure, and not degrading the time delay margins of the nominal design.

DF-MRAC is particularly well suited for these circumstances.

Quantification of transient dynamics, guaranteed transient and steady-state per-

formance bounds, and gain/time-delay sensitivity at the system inputs and outputs

are three open problems in adaptive control theory. An analysis based on Singu-

lar Perturbation Theory is used to quantify the transient dynamics of an adaptively

controlled system in Ref. 52. A similar analysis can be applied to derivative-free

adaptive control which should provide a similar result. With regard to the second

problem, we show that DF-MRAC does have guaranteed transient and steady-state
performance bounds. In order to reduce gain/time-delay sensitivity, we combined

derivative-free adaptive control with ALR-modification as proposed in Ref. 10. It is

shown in both simulation using NASA’s generic transport model (GTM) [1] and in

flight test on NASA Airborne Subscale Transport Aircraft Research (AirSTAR) flight

test vehicle [41, 42] that DF-MRAC with ALR-modification is successful in dealing

with uncertainties and latencies with the loop broken at the plant input.

6
1.5 Output Feedback Adaptive Control

The assumption of full state feedback leads to computationally simpler adaptive con-

troller algorithms in comparison to output feedback algorithms. Output feedback

adaptive controllers, however, are required for applications in which it is impractical

or impossible to sense the entire state of the process, such as active noise suppression,

active control of flexible structures, fluid flow control systems, combustion control

processes, and low cost or expendable unmanned aerial vehicle applications. Models

for these applications vary from reasonably accurate low-frequency models, in the

case of structural control problems, to less accurate low-order models in the case of
active control of noise, vibrations, flows, and combustion processes.

There has been a number of results in the recent decades focused on output feed-

back adaptive controllers. For example, Esfandiari and Khalil [27] introduced a high

gain observer for the reconstruction of the unavailable states. Krstic, Kanellakopou-

los, and Kokotovic [49], and Marino and Tomei [56] proposed backstepping approaches

for output feedback adaptive control, which are affine with respect to unknown pa-
rameters. In Ref. 48, Kim and Lewis suggested a neural network architecture for the

observer. Output feedback adaptive control using a high gain observer and neural

networks has also been proposed in Ref. 69 by Seshagiri and Khalil for nonlinear

systems represented by input output models. Another method involving the design

of adaptive observer using function approximators and backsteppting was given by

Choi and Farrel [18] for a limited class of systems that can be transformed to output
feedback form in which nonlinearities depend on measurements only.

Calise, Hovakimyan, and Idan [11] developed an output feedback adaptive control

architecture for nonlinear uncertain dynamical systems without using a state observer

but with a stable low-pass filter to satisfy a strictly positive real conditions. It requires

that the input vector to the neural networks be composed of current and past input

output data. In Ref. 34, Hovakimyan, Nardi, Calise, and Kim considered an output

7
feedback adaptive controller with an error observer instead of a state observer. In
this approach, only the relative degree of the regulated system needs to be known

and the basis functions employed depend only the available input output history.

Volyanskyy, Haddad, and Calise [74] introduced Q-modification to output feedback

adaptive control. This approach assumes a restrictive matching assumption.

Applying an adaptive control technique for uncertain dynamical systems in many

cases implies replacement of an existing control system. On the other hand, it is

highly desirable to consider an adaptive approach that can be implemented in a form

of augmenting an existing controller. This rationale has been one of the viewpoints

taken in applying adaptive control in Refs. 2, 3, 6, 21, 80.

Hoagg, Santillo, and Bernstein [32] estimate Markov parameters that leads to out-

put feedback adaptive control algorithms with varying set of assumptions. Recently,

Lavretsky [51] introduced an adaptive output feedback design using asymptotic prop-
erties of LQG/LTR controllers that asymptotically satisfies a strictly positive real

(SPR) condition. In this approach, it is assumed that the determinant of the product

of the system output and input matrices is not zero, although it is possible to relax

this assumption.

In this dissertation, we extend the derivative-free adaptive control architecture

mentioned in the previous section to output feedback for adaptively controlling con-
tinuous time uncertain dynamical systems using a state observer. The observer is em-

ployed in the adaptive part of the design in place of a reference model. A derivative-

free weight update law in output feedback form ensures that the estimated states

follow both the reference model states and the true states so that both the state

estimation error and the state tracking error are bounded. The proposed approach

is particularly advantageous for applications to systems that can undergo a sudden

change in dynamics, such as might be due to reconfiguration, deployment of a payload,

docking, or structural damage.

8
 The stability analysis uses a Lyapunov-Krasovskii functional that entails the solu-
tion of a parameter dependent Riccati equation [46], rather than a Lyapunov equation,

to show that all the error signals are uniformly ultimately bounded (UUB). The com-

plexity of the proposed approach is significantly simpler than many other approaches

to output feedback adaptive control, and it can be implemented in a form that aug-

ments an existing observer based linear control architecture. The design procedure is

illustrated using an aeroelastic generic transport model.

1.6 Organization

The organization of this dissertation is as follows. Chapter II presents the theory of

KF based modification in adaptive control. The results of this chapter are extended in

Chapter III to obtain a KF based adaptation law. Chapter IV presents the theory of

DF-MRAC. The results of this chapter are extended in Chapter V to output feedback

adaptive control. Finally, Chapter VI concludes this dissertation.

1.7 Notation

The notation used in this dissertation is fairly standard. IRn denotes the set of

n × 1 real column vectors, IRn×m denotes the set of n × m real matrices, (·)T denotes

transpose, (·)−1 denotes inverse, ai,j ∈ A denotes the row i, column j element of

matrix A, and S ⊂ R denotes S being a subset of R. Furthermore, we write λmin(M)

(resp., λmax (M)) for the minimum (resp., maximum) eigenvalue of the Hermitian

matrix M, vec(·) for the column stacking operator, A ⊗ B for the Kronecker product,
diag[A, B] for a block diagonal matrix formed with matrices A and B on the diagonal,

| · | for the Euclidean vector norm, || · || for the Euclidean matrix norm, tr(·) for trace

operator, R{·} denotes the range space of a matrix, (a, b) for the open interval in IR

from a to b > a, and [a, b] for the closed interval in IR from a to b > a.

9
 CHAPTER II

Kalman Filter Modification in Adaptive Control

2.1 Introduction

This chapter presents a novel KF based approach for approximately enforcing a linear

constraint in adaptive control which leads to an alternative forms for many commonly
used modification terms. Simulation results are provided that illustrate that this

approach provides smaller tracking errors without incurring significant oscillations in

the system response and without requiring a high modification gain.

For illustrative purposes, we develop a KF version of the well known e–modification.

It is shown how the standard e–modification term [37] can be interpreted as the gradi-

ent of a norm measure of a linear constraint violation, and then this linear constraint is
used to develop a Kalman filter based e–modification. As a second example, the linear

constraint imposed by ALR–modification [10] is treated using the KF method. The re-

sulting KF-ALR–modification is shown to improve upon standard ALR–modification.

Finally, a solution is proposed to the problem of adaptation in the presence of input

constraints that does not require modification of the reference model.

2.2 Preliminaries

We begin by presenting a formulation of the model reference adaptive control problem.

For this purpose, consider the following uncertain system

 
ẋ(t) = Ax(t) + B u(t) + ∆(x(t)) (1)

10
where x(t) ∈ Rn is the state vector, u(t) ∈ Rm is the control input, A ∈ Rn×n and
B ∈ Rn×m are known matrices with (A, B) being a controllable pair, and ∆(·) :

Rn → Rm is the unknown matched uncertainty. Furthermore, it is assumed that

x(t) is available for feedback and u(t) is restricted to the class of admissible controls

consisting of measurable functions.

When controlling a nonlinear system, the matrices A, B in (1) are usually obtained

by linearizing the nonlinear dynamics at selected equilibrium conditions, and the

resulting set of linear models are used to design a gain scheduled controller. In this

setting a part of the uncertainty in (1) can be thought of as arising from the modeling

error associated with linearization. It is assumed that a nominal (baseline) controller

for the system in (1) exists for a neighborhood of each equilibrium point, and can be

written in the form

un (t) = −K1 x(t) + K2 r(t) (2)

where r(t) ∈ Rr , r ≤ m, is a bounded reference command that defines the desired

output response, K1 ∈ Rm×n is the state feedback gain matrix, and K2 ∈ Rm×r is
the command input gain matrix. It should be noted that many controller designs

commonly contain dynamics, in which case one can augment the controller dynamics

with the dynamics in (1), and consider an expanded state made up of the system

states and the controller states, and rewrite the dynamics and controller in the form

of (1) and (2). So there is no loss in generality with respect to dynamic controllers in

assuming these forms. However, to further simplify the discussion we introduce the
following assumption.

Assumption 2.1. The matched uncertainty in (1) can be linearly parameterized

as

∆(x) = W T β(x) (3)

where W ∈ IRs×m is the unknown constant weight matrix and β : IRn → IRs is a

11
known vector of basis functions of the form β(x) = [β1 (x), β2 (x), . . . , βs (x)]T ∈ IRs .
One can construct a reference model for the desired response of the closed loop

system as

ẋm (t) = Am xm (t) + Bm r(t) (4)

where xm (t) ∈ Rn is the reference state vector, Am ∈ IRn×n is Hurwitz, and Bm ∈

IRn×r .

Assumption 2.2. Am and Bm in (4) are chosen so that:

Am = A − BK1

Bm = BK2

Consider the augmenting adaptive controller given by

u(t) = un (t) − uad (t) (5)

uad (t) = Ŵ (t)T β(x(t)) (6)

Defining

e(t) ≡ x(t) − xm (t)

it is well known [71] that the following adaptive law

˙  
Ŵ (t) = γ β(x(t))eT (t)P B (7)

where γ is a positive adaptation gain (adaptation learning rate), and P ∈ Rn×n is a

positive definite solution of the Lyapunov equation

0 = AT
m P + P Am + Qm (8)

for any Qm > 0, ensures that Ŵ (t) remains bounded, and that e(t) → 0 as t → ∞.

The resulting adaptive control system is illustrated in Figure 1.

12
 Figure 1: Augmenting adaptive control of a baseline control system.

Definition 2.1 ([30,45]). Consider the nonautonomous system ẋ = f (t, x), where

f : [0, ∞) × D → IRn is piecewise continuous in t, locally Lipschitz in x on [0, ∞) × D,

and D ⊂ IRn is a domain that contains the origin, x = 0. If there exists positive

constant r and r̄, independent of t0 ≥ 0, and for every η ∈ (0, r̄), there is T = T (η, r),

independent of t0 ≥ 0, such that ||x(t0 )|| ≤ η ⇒ ||x(t)|| ≤ r for all t ≥ t0 +T , then the

solution of ẋ = f (t, x) is uniformly ultimately bounded (UUB) with ultimate bound

r.

In the case where there are bounded external disturbances or when Assumption
2.1 is relaxed to the following assumption, it can be shown that e(t) is UUB.

Assumption 2.3. The matched uncertainty in (1) can be linearly parameterized

as

∆(x) = W T β(x) + ε(x), |ε(x)| ≤ ǫ, x ∈ Dx , Dx ⊂ IRn (9)

where W ∈ IRs×m is the unknown constant weight matrix, β : IRn → IRs is a

13
known vector of basis functions of the form β(x) = [β1 (x), β2 (x), . . . , βs (x)]T ∈ IRs ,
ε(x) : IRn → IRm is the residual error, and Dx is a sufficiently large compact set.

In this case the adaptive law in (7) can be modified to guarantee that Ŵ (t) remains

bounded, using either σ–modification [37], e–modification [60], and/or parameter

projection [66]. For example with e–modification the adaptation law becomes

˙  
Ŵ (t) = γ β(x(t))eT (t)P B − σ|e(t)|Ŵ (t) (10)

where σ is a positive modification gain (modification learning rate).

The state error dynamics are obtained using (1) and (4) with (2), (5), (6), and

(10)

 
ė(t) = Ax(t) + B u(t) + W T β(x) + ε(x) −Am xm (t) − Bm r(t)
 
= Ax(t) + B −K1 x(t) + K2 r(t) − Ŵ (t)T β(x(t)) + W T β(x(t)) + ε(x)

−Am xm (t) − Bm r(t)

= Am x(t) + Bm r(t) + B W̃ T (t)β(x(t)) + Bε(x(t)) − Am xm (t) − Bm r(t)

= Am e(t) + B W̃ T (t)β(x(t)) + Bε(x(t)) (11)

where W̃ (t) ≡ W − Ŵ (t). Since W is a constant matrix, it follows from (10) that

˙ (t) = −γ β(x(t))eT (t)P B − σ|e(t)|Ŵ (t)

W̃ (12)

Define ζ(t) = [e(t)T vec(W̃ (t))T ]T and Br = {ζ(t) : |ζ(t)| ≤ rα }, such that Br ⊂ Dx

for a sufficiently large compact set Dx . Denote Ωα = {ζ(t) ∈ Br : ζ(t)T P̃ ζ(t) ≤ α},
min
where P̃ = diag[P, γ −1 I] and α = |ζ(t)|=rα
(ζ(t)T P̃ ζ(t)) = rα2 λmin (P̃ ). A depiction of

these sets is given in Figure 2. Further define Θe = ( σ2 ||W ||2 + 2||P B||ε)/λmin(Qm ),
p λmax (P )Θ2e + γ1 Θ2
ΘW̃ = 12 ||W ||2 + √1σ ||P B||ε + σ4 ||W ||2, and r 2e = λ (P̃ )
W̃
. We can then
min

state the following theorem.

Theorem 2.1. Consider the uncertain system given by (1) subject to Assumption

2.3. Consider, in addition, the feedback control law given by (5), with the nominal

14
 Figure 2: Geometric representation of sets.

feedback control component given by (2), and with the adaptive feedback control

component given by (6) that has a weight update law in the form (10). If ζ(t0) ∈ Ωα

and rα > r e , then either ζ(t) is UUB with ultimate bound r e , or e(t) = 0 for all
t ≥ T ≥ t0 and W̃ (t) is bounded.

Proof. Consider the Lyapunov-like function candidate:

1 T 1 1
V(e(t), W̃ (t)) = e (t)P e(t) + tr[W̃ T (t)W̃ (t)] = ζ T (t)P̃ ζ(t) (13)
2 2γ 2

The time derivative of (13) along the trajectories of (11) and (12) can be expressed

as

1 ˙
V̇(e(t), W̃ (t)) = eT (t)P [Am e(t) + B W̃ T (t)β(x(t)) + Bε(x(t))] − tr[W̃ T (t)Ŵ (t)]
γ
1 T
= − e (t)Qm e(t) + eT (t)P Bε(x(t)) + σtr[W̃ T (t)|e(t)|(W − W̃ (t))]
2
 
≤ −|e(t)| c1 |e(t)| + c2 ||W̃ (t)|| − c3 − c4 ||W̃ (t)||
 √ c4
2 c2 
= −|e(t)| c1 |e(t)| + c2 ||W̃ (t)|| − √ −c3 − 4 (14)
2 c2 4c2

where c1 = λmin (Qm )/2, c2 = σ, c3 = ||P B||ε, and c4 = σ||W ||. Note that

V̇(e(t), W̃ (t)) is negative as long as the term in braces is positive and e(t) 6= 0.

15
 ˙ (t) = 0 from (12). For the
If e(t) = 0 for all t ≥ T , then W̃ (t) is bounded since W̃
case when e(t) 6= 0, either |e(t)| > Θe , or ||W̃ (t)|| > ΘW̃ renders V̇ (e(t), W̃ (t)) < 0,
q
1 c24
1 c c2

where Θe = c1 c3 + 4c2 , and ΘW̃ = √c2 2√c2 + c3 + 4c42 . Therefore, e(t) and W̃ (t)
4

are uniformly bounded. Similar to [12], using e(t)T P e(t) ≤ λmax (P )|e(t)|2 , (14) can

be written as
p p
e(t)T P e(t)  e(t)T P e(t) √ c4
2 c2 
V̇(e(t), W̃ (t)) ≤ − p c1 p + c2 ||W̃ (t)|| − √ −c3 − 4
λmax (P ) λmax (P ) 2 c2 4c2
p p
which is strictly negative when e(t)T P e(t) > λmax (P )Θe or ||W̃ (t)|| > ΘW̃ .

Define Ωβ = {ζ(t) ∈ Br : ζ(t)T P̃ ζ(t) ≤ λmax (P )Θ2e + γ1 Θ2W̃ }. Then, Ωα is a positively

invariant set if Ωβ ⊂ Ωα . This requires that λmax (P )Θ2e + γ1 Θ2W̃ < α. The minimum
size of Br that ensures this condition is rα > re . Therefore, if ζ(t0 ) ∈ Ωα , then error

dynamics (11) and (12) are UUB with ultimate bound re . 

Remark 2.1. In Theorem 2.1 it should be noted that ζ(t) is UUB with probability

one in the sense that given a joint probability density function for the set of initial

conditions {e(t0 ), W̃ (t0 )}, the probability measure of all initial conditions from that

set that will result in e(t) = 0 for t ≥ T is zero, because the space of initial conditions
that produces this outcome has dimension less than n + m × s.

Remark 2.2. When the uncertainty is exactly parameterized, e-modification is

not needed to prove the weight error remains bounded. In this case c2 = 0 and c3 = 0,

and V̇(e(t), W̃ (t)) = −eT (t)Qm e(t) ≤ −c1 |e(t)|2 , which is sufficient to prove that e(t)

and W̃ (t) are bounded. Since V(e(t), W̃ (t)) is lower bounded and V̇(e(t), W̃ (t)) ≤

0, V(e(t), W̃ (t)) approaches a finite limit as t → ∞. Moreover, V̇(e(t), W̃ (t)) is

a uniformly continuous function of time, since its time derivative V̈(e(t), W̃ (t)) =
 

−eT (t)Qm Am e(t) + B W̃ T (t)β(x(t)) is bounded, which is a consequence of the fact

that e(t) and W̃ (t) are bounded. Then by Barbalat’s lemma [30] it follows that

limt→∞ V̇(e(t), W̃ (t)) = 0, and therefore e(t) asymptotically goes to zero.

Remark 2.3. The condition rα > r e is needed to ensure that x(t0 ) lies inside the

16
ball Dx as defined in Assumption 2.3. Invariancy in the error dynamics together with
the fact that xm (t) is bounded ensures that x(t) remains in Dx . It is apparent from

this condition that this implies a lower bound on γ whenever σ > 0. However, it also

implies an upper bound. When γ is sufficiently large, then λmin (P̃ ) = 1/γ and the

limiting from of this condition r 2e > γλmax (P )Θ2e .

In the next section, we introduce an optimization based approach for approxi-

mately imposing a linear constraint.

2.3 Approximate Enforcement of a Linear Constraint by

KF Optimization

We first restrict the form of the constraint of interest. For this purpose, the constraint

on the weight estimate in an adaptive control design has the linear form

Ŵ T (t)φ(t, x(t), e(t)) = 0 (15)

where φ(·) : [0, ∞) × IRn × IRn → IRs×l is known and bounded pointwise in time,

with l ≤ s.

The problem of estimating W while approximately enforcing the linear constraint

in (15) can be formulated as an optimization problem. Eq. (15) can be expressed in

the following equivalent vector form

vec φT (t, x(t), e(t))Ŵ (t) = Φ(t, x(t), e(t))T ω = 0 (16)

where Φ(·) = Im×m ⊗ φ(·) : [0, ∞) × IRn × IRn → IRms×ml and ω = vec(W ) ∈ IRms .

Define the stochastic process

ω̇ = q(t) (17)

z(t) = ΦT (·)ω + r(t) (18)

where q(t) and r(t) are zero-mean, Gaussian, white noise processes with covariances

E q(t)q T (τ ) = Q̄δ(t − τ ), Q̄ ∈ IRms×ms > 0

E r(t)r T (τ ) = R̄δ(t − τ ), R̄ ∈ IRml×ml > 0

17
and z(t) is regarded as a measurement. The estimate of z(t) is

ẑ(t) = ΦT (·)ω̂(t)

where ω̂(t) is an estimate of ω. The Kalman filter associated with this problem
formulation is given by [19]:

 
˙
ω̂(t) = S̄(t)Φ(·)R̄−1 z(t) − ẑ(t) , ω̂(0) = 0 (19)
˙
S̄(t) = −S̄(t)Φ(·)R̄−1 ΦT (·)S̄(t) + Q̄, S̄(0) = S̄0 > 0 (20)

where S̄(t) ∈ IRms×ms . Since the objective is to approximately satisfy the constraint

in (15), the logical choice for z(t) when employing this estimator is z(t) = 0.

Lemma 2.1. Let R̄ = Im×m ⊗ R, with R ∈ IRl×l > 0, and Q̄ = Im×m ⊗ Q, with

Q ∈ IRs×s > 0. Then, Eqs. (19) and (20), with z(t) = 0, reduces to:

˙
Ŵ (t) = −S(t)φ(·)R−1 φT (·)Ŵ (t), Ŵ (0) = 0 (21)

Ṡ(t) = −S(t)φ(·)R−1 φT (·)S(t) + Q, S(0) = S0 > 0 (22)

where S(t) ∈ IRs×s .

Proof. Substituting z(t) = 0, (19) becomes

˙
ω̂(t) = −S̄(t)Φ(·)R̄−1 ΦT (·)ω̂

Choosing R̄ = Im×m ⊗ R and Q̄ = Im×m ⊗ Q, S̄(t) in (20) will have the form of

S̄(t) = Im×m ⊗ S(t), with S(t) ∈ IRs×s . Now, it follows from the above equation with

18
ω̂ = vec(Ŵ (t)) ∈ IRms and Φ(·) = Im×m ⊗ φ(·) ∈ IRms×ml that
˙
vec(Ŵ ) = −[Im×m ⊗ S(t)][Im×m ⊗ φ(·)][Im×m ⊗ R−1 ]

×[Im×m ⊗ φT (·)]vec(Ŵ (t))

= −[Im×m ⊗ S(t)][Im×m ⊗ φ(·)][Im×m ⊗ R−1 ]

×vec(φT (·)Ŵ (t))

= −[Im×m ⊗ S(t)][Im×m ⊗ φ(·)]vec(R−1 φT (·)Ŵ (t))

= −[Im×m ⊗ S(t)]vec(φ(·)R−1φT (·)Ŵ (t))

= −vec(S(t)φ(·)R−1φT (·)Ŵ (t))

where the fact (C T ⊗ A)vec(B) = vec(ABC) is used [5], with C = Im×m . It follows

from vec−1 operator that the above equation becomes (21). Similarly, it follows

from(20) that

[Im×m ⊗ Ṡ(t)] = −[Im×m ⊗ S(t)][Im×m ⊗ φ(·)][Im×m ⊗ R−1 ]

×[Im×m ⊗ φT (·)][Im×m ⊗ S(t)] + [Im×m ⊗ Q]

= −[Im×m ⊗ S(t)φ(·)R−1φT (·)S(t)]

+[Im×m ⊗ Q]

where the fact (A ⊗ B)(C ⊗ D) = (AC ⊗ BD) is used [5], with A = C = Im×m . Since,

[I ⊗ A] = [I ⊗ B] + [I ⊗ C] implies A = B + C, then the above equation becomes

(22). 

Remark 2.4. The Riccati equation in (22) can be associated with the following

system of equations:

ẇi = q1 (t)

zi (t) = φT (·)wi + r1 (t)

in which wi are the process dynamics associated with the i-th column of W , zi is the
 
corresponding measurement vector, E q1 (t)q1T (τ ) = Qδ(t − τ ), and E r1 (t)r1T (τ ) =

Rδ(t − τ ).

19
 Assumption 2.4. The observability gramian
Z t
W0 (t0 , t) = φ(t, x(t), e(t))φT (t, x(t), e(t))dt (23)
t0

of the system defined in Remark 2.4 satisfies the following condition for uniform

complete observability [43]:

0 < α0 (σ)Il ≤ W0 (t0 , t0 + σ) ≤ α1 (σ)Il (24)

for all t0 , and for some fixed constant σ > 0.

Remark 2.5. The condition in Assumption 2.4 is sufficient to show that S(t) is

positive-definite, and lower and upper bounded, independent of t0 , if φ(t, x(t), e(t)) is

bounded pointwise in time. That is, there exists constants s > 0 and s̄ > 0 such that

sIs ≤ S(t) ≤ s̄Is holds for all t ≥ t0 .

Using Lemma 2.1 and incorporating (21) as a modification term in the adaptive

law given in (10), we arrive at

˙
Ŵ (t) = γ(β(x(t))eT (t)P B − σ|e(t)|Ŵ (t) − kS(t)φ(·)R−1 φ(·)T Ŵ (t)) (25)

where k has been introduced a positive modification gain, and kS(t) can be interpreted

as its variable gain. In this case the weight update error dynamics become

˙ (t) = −γ(β(x(t))eT (t)P B − σ|e(t)|Ŵ (t) − kS(t)φ(·)R−1 φ(·)T Ŵ (t))

W̃

A stability analysis based on the above weight update error dynamics depends on the

form of the constraint in (15), and is therefore problem dependent. For illustration

purposes, we state and prove the KF alternative to e–modification in the next section.

2.4 Examples on a Model of Wing Rock Dynamics

In this section we first show how the KF approach for enforcing a linear constraint

can be used to derive an alternative for the standard e–modification term in adaptive

20
control. Next an alternative to ALR–modification is illustrated [10]. As a third
example, we introduce a novel solution to the problem of input constraints imposed

by an actuator, and compare it with the method of hedging [39, 40]. All of the

examples are treated using an adaptive control approach to the problem of stabilizing

uncertain wing rock dynamics [70]. In all cases we chose Qm = I2 , and for the KF

based designs we chose Q = I3 , R = 10−3 I3 , and S0 = I3 . The form in Assumption

2.3 was used to approximate the uncertainty with β(x(t)) = [1, β1 (x(t)), β2 (x(t))]T ,
where βi (x) = 1/(1 + e−xi ), i = 1, 2, are sigmoidal type basis functions.

2.4.1 A KF Alternative to e–Modification

In this subsection we replace the standard e–modification with KF e–modification.

The standard e–modification term in (10) can be viewed as having been introduced

to approximately enforce the following linear constraint

1
|e(t)| 2 Ŵ (t) = 0 (26)

˙ 1
by adding a component to Ŵ (t) along the gradient of || |e(t)| 2 Ŵ (t) ||2F with respect
1
to Ŵ (t). In this case φ(·) = |e(t)| 2 Is in (15). In this case it will be necessary to prove

that φ(|e|) remains bounded pointwise in time as an intermediate step in the proof.

Applying the form of constraint in (26) to (21) and (22) leads to the following KF

based e–modification term:

˙
Ŵ (t) = −|e(t)|S(t)R−1 Ŵ (t)

Ṡ(t) = −|e(t)|S(t)R−1 S(t) + Q

Remark 2.6. If Q, R, and S0 are chosen in the form q̄I, r̄I, and s̄o I, respectively,

then the above equations reduce to:

˙ |e(t)|s(t)
Ŵ (t) = − Ŵ (t) (27)
r̄
s2 (t)
ṡ(t) = −|e(t)| + q̄ (28)
r̄

21
 Replacing the conventional e–modification term in Eq. (10) with the KF based
term in Eq. (27) we have
h |e(t)|s(t) i
˙
Ŵ (t) = γ β(x(t))eT (t)P B − σkf Ŵ (t) (29)
r̄

where σkf is a positive learning rate and s(t) is defined in (28). In this case, the

weight update error dynamics can be given using (29) as:

h i
˙ (t) = −γ β(x(t))eT (t)P B − σ |e(t)|s(t) Ŵ (t)
W̃ (30)
kf
r̄
1
Furthermore, note that since φ(·) = |e(t)| 2 Is , it follows from (23) that W0 (t0 , t) =
Rt
t0
|e(τ )|dτ Is is positive-definite for all t ≥ t0 as long as |e(τ )| =
6 0 over the interval

[t0 , t].
1 2
λmax (P )Θ̄2e + Θ̄ σkf s̄
Define r 2ekf = λmin (P̃ )
γ W̃
, where Θ̄e = ( 2r̄
||W ||2 + 2||P B||ǫ)/λmin(Qm ), Θ̄W̃
√ q
σkf s̄
= 21 ||W ||2 + √σkfr̄ s ||P B||ǫ + 4r̄
||W ||2, where s and s̄ are defined in Remark 2.5.

Theorem 2.2. Consider the uncertain system given by (1) subject to Assumption

2.3. Consider, in addition, the feedback control law given by (5), with the nominal

feedback control component given by (2), and with the adaptive feedback control

component given by (6) that has a weight update law in the form (29). If ζ(t0) ∈ Ωα

and rα > rekf , then either ζ(t) is UUB with ultimate bound r ekf , or e(t) = 0 for all
t ≥ T ≥ t0 and W̃ (t) is bounded.

Proof. Consider the Lyapunov-like function candidate given by (13). The time

derivative of (13) along the trajectories of (11) and (30) can be expressed as:

1 ˙
V̇(e(t), W̃ (t)) = eT (t)P [Am e(t) + B W̃ T (t)β(x(t)) + Bε(x(t))] − tr[W̃ T (t)Ŵ (t)]
γ
1
= − eT (t)Qm e(t) + eT (t)P Bε(x(t)) + σkf s(t)tr[W̃ T (t)
2
·|e(t)|(W − W̃ (t))]/r̄
 
≤ −|e(t)| d1 |e(t)| + d2 (t)||W̃ (t)|| − d3 − d4 (t)||W̃ (t)||
 p d4 (t)
2 d2 (t) 
= −|e(t)| d1 |e(t)| + d2 (t)||W̃ (t)|| − p −d3 − 4 (31)
2 d2 (t) 4d2 (t)

22
where d1 = λmin(Qm )/2, d2 (t) = σkf s(t)/r̄, d3 = ||P B||ε, and d4 (t) = σkf s(t)||W ||/r̄.
Note that V̇(e(t), W̃ (t)) is negative as long as the term in braces is positive and

e(t) 6= 0. If e(t) = 0 for all t ≥ T , then W̃ (t) is bounded since W̃ ˙ (t) = 0 from (30).
d2 (t)
d4 (t)
For the case when e(t) 6= 0, either |e(t)| > d11 d3 + 4d42 (t) or ||W̃ (t)|| > √ 1 √ +
d2 (t) 2 d2 (t)
q
d2 (t)

d3 + 4d42 (t) renders V̇ (e(t), W̃ (t)) < 0. At t = t0 , s(t0 ) > 0. Since both e(t) and

s(t) are continuous functions of time, there exists a finite time t1 > t0 such that both

s(t) and e(t) are bounded. During this time interval s(t) > 0, since s(t) is continuous

and must pass through 0 in order to become negative, at which point ṡ(t) = q̄ > 0.

In addition, during this time interval ζ(t) is confined to the largest invariant set, Ωα ,
contained in Br , such that r > r ekf , since we are assured that V̇ (e(t), W̃ (t)) < 0

outside the ball Br e . By repeating this argument for some ti > ti−1 , i = 2, 3, . . .,
kf

one can extend the time interval indefinitely beyond t1 , to show that φ(|e|) remains

bounded pointwise in time. Therefore, by Remark 2.5 s(t) satisfies s ≤ s(t) ≤ s̄ for

all t ≥ t0 .

Next, let d2 = σkf s/r̄, d¯4 = σkf s̄||W ||/r̄, and note from the previous definitions
q
1 d¯24
1 ¯4
d d¯24

Θ̄e and Θ̄W̃ , that Θ̄e = d1 d3 + 4d , and Θ̄W̃ = √ √ + d3 + 4d . Since
2 d2 2 d2 2
q
d2 (t)
d4 (t) d2 (t)

Θ̄e > d11 d3 + 4d42 (t) and Θ̄W̃ > √ 1 √ + d3 + 4d42 (t) , V̇ (e(t), W̃ (t)) is negative
d2 (t) 2 d2 (t)

for either |e(t)| > Θ̄e or ||W̃ (t)|| > Θ̄W̃ , independent of t. Similar to the proof of

Theorem 2.1, define Ωβ = {ζ(t) ∈ Br : ζ(t)T P̃ ζ(t) ≤ λmax (P )Θ̄2e + γ1 Θ̄2W̃ }. Then, Ωα is

a positively invariant set if Ωβ ⊂ Ωα . This requires that λmax (P )Θ̄2e + γ1 Θ̄2W̃ < α. The

minimum size of Br that ensures this condition is rα > r ekf . Therefore, if ζ(t0) ∈ Ωα ,

then the error dynamics (11) and (12) are UUB with ultimate bound r ekf . 

Next, we compare the standard e–modification based adaptation law in Eq. (10)
with the KF version of e–modification in Eq. (29) using a model of wing rock dynamics

to illustrate the advantages of the proposed KF approach to modification.

23
 For this purpose, consider the following dynamics
      
ẋ1 (t) 0 1 x1 (t) 0  
 =   +   u(t) + ∆(x(t)) (32)
ẋ2 (t) 0 0 x2 (t) 1

where ∆(x(t)) = α1 x1 (t) + α2 x2 (t) + α3 |x1 (t)|x2 (t) + α4 |x2 (t)|x2 (t) + α5 x31 (t) with

α1 = 0.2314, α2 = 0.6918, α3 = −0.6245, α4 = 0.0095, and α5 = 0.0214. In (32),

x1 (t) represents the roll angle, and x2 (t) represents the roll rate. We selected the initial

state values as x(0) = [6o , 3o /s]T . The reference system is selected to be second order

with a natural frequency of 0.5 rad/s, and a damping of 0.707, and to have unity low

frequency gain from r(t) to x1 (t). This corresponds to choosing K1 = [0.25, 0.707]

and K2 = 0.25. Figures 3-10 present the results.

Figure 3 shows that the baseline control response with uncertainty significantly de-
grades system performance. Figures 4-8 compare the standard e–modification based

adaptive control law with the KF based e–modification adaptive control law for dif-

ferent gains. It is obvious from these figures that performance is good only when

we employ KF based e–modification. The standard e–modification based adaptive

control law does not produce a reasonable result for any of these cases. In addition,

KF based e–modification requires significantly less and smoother control effort. The
gain in the KF based approach to e–modification is now time varying due to the

presence of S(t), and that its performance is nearly independent of value of gains

employed. Since most actuators are rate limited, in Figure 8 we compare du(t)/dt

responses. Finally, Figure 9 and 10 show the comparison results under sensor noise,

where it is evident that the KF based e–modification achieved the best performance

while requiring much less control effort.

At this point the reader may wonder if this comparison of e–modification is biased

by the fact that no attempt was made to tune the adaptive law other than to vary

the adaptation gains. However this is not the case since all our attempts to tune

the adaptive law with conventional e–modification did not significantly change the

24
 20 3

without uncertainty
15
with uncertainty 2

x2 [degree/s]
10

x1 [degree]
1
5
0
0

−1
−5

−10 −2
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time [s] Time [s]

0.04 1

0.02
0.5
Control Signal

|| W ||
−0.02 0

−0.04
−0.5
−0.06

−0.08 −1
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time [s] Time [s]

Figure 3: Baseline control response without and with uncertainty.

outcome [13].

25
 30 80

20 60

10 40

x2 [degree/s]
x1 [degree]
0 20

−10 0

−20 −20

−30 −40

−40 −60
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 4

3.5 Eq. (10), γ=1, σ=0

2 Eq. (10), γ=1, σ=1
3 Eq. (29), γ=1, σkf=1
Control Signal

1
2.5

||W||
0 2

1.5
−1
1
−2
0.5

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 4: Comparison of adaptive controller responses without modification, with

e-modification, and with KF based e- modification with all gains set to 1.

8 15

Eq. (10), γ=25, σ=0

10
6 Eq. (10), γ=25, σ=25
Eq. (29), γ=25, σkf=25
x2 [degree/s]

5
x1 [degree]

4
0
2
−5

0
−10

−2 −15
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 5

2
4
Control Signal

1
3
||W||

0
2
−1

1
−2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 5: Comparison of adaptive controller responses without modification, with

e-modification, and with KF based e- modification with all gains set to 25.

26
 8 8
Eq. (10), γ=100, σ=0
6
6 Eq. (10), γ=100, σ=100
Eq. (29), γ=100, σkf=100 4

x2 [degree/s]
x1 [degree]
4 2

0
2 −2

−4
0
−6

−2 −8
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 5

2
4
Control Signal

1
3

||W||
0
2
−1

1
−2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 6: Comparison of adaptive controller responses without modification, with

e-modification, and with KF based e- modification with all gains set to 100.

8 5
Eq. (10), γ=250, σ=0
Eq. (10), γ=250, σ=250
6
Eq. (29), γ=250, σkf=250
x2 [degree/s]
x1 [degree]

4
0
2

−2 −5
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 5

2
4
Control Signal

1
3
||W||

0
2
−1

1
−2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 7: Comparison of adaptive controller responses without modification, with

e-modification, and with KF based e- modification with all gains set to 250.

27
 100 Eq. (10), γ=250, σ=0
Eq. (10), γ=250, σ=250
Eq. (29), γ=250, σkf=250
80

60

40

20
du/dt

−20

−40

−60

−80

−100
0 5 10 15 20 25 30 35 40
Time [s]

Figure 8: Comparison of du(t)/dt responses without modification, with e-

modification, and with KF based e- modification with all gains set to 250.

10 20
Eq. (10), γ=25, σ=0 15
8
Eq. (10), γ=25, σ=25
Eq. (29), γ=25, σkf=25 10
6
x2 [degree/s]
x1 [degree]

5
4
0
2
−5
0
−10
−2 −15

−4 −20
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

4 7

3 6
2
5
Control Signal

1
4
||W||

0
3
−1
2
−2

−3 1

−4 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 9: Comparison of adaptive controller responses under sensor noise without

modification, with e-modification, and with KF based e- modification with all gains
set to 25.

28
 40 Eq. (10), γ=25, σ=0
Eq. (10), γ=25, σ=25
Eq. (29), γ=25, σkf=25

20

0
du/dt

−20

−40

−60

−80
0 5 10 15 20 25 30 35 40
Time [s]

Figure 10: Comparison of du(t)/dt responses under sensor noise without modifica-
tion, with e-modification, and with KF based e- modification with all gains set to
25.

29
2.4.2 A KF Alternative to ALR–Modification

In Ref. [10], an ALR–modification term is proposed for use in adaptive control to

preserve the stability margins of the reference model defined in (4). This is done by

approximately imposing a linear constraint on the weights so that the loop properties

of the reference model are asymptotically recovered as the gain on the modification

term is increased. Here, the proposed KF optimization approach is applied to the ALR
dβ(x(t))
constraint Ŵ T (t)βx (x(t)) = 0 as a second example, where βx (x(t)) ≡ dx(t)
∈ Rs×n
is the derivative of the basis function with respect to x(t) ∈ Rn . The standard ALR–

modification term is found by taking the negative gradient of the following quadratic

function in Ŵ (t):

1
JA (Ŵ (t)) = tr[(Ŵ T (t)βx (x(t)))T (Ŵ T (t)βx (x(t)))]
2
∂JA (Ŵ (t))
= βx (x(t))βxT (x(t))Ŵ (t)
∂ Ŵ (t)
Introducing the above gradient as a modification to adaptive law in (10) results in

˙  
Ŵ (t) = γ β(x(t))eT (t)P B − σ|e(t)|Ŵ (t) − kgrad βx (x(t))βxT (x(t))Ŵ (t) (33)

where kgrad is the gradient learning rate. Since φ(·) = βx (x(t)) for the ALR–

modification, using this expression in (21) and (22), and applying the result as a

modification to the adaptive law in (10) leads to the following adaptive law

˙ 
Ŵ (t) = γ β(x(t))eT (t)P B − σ|e(t)|Ŵ (t)

−kkf S(t)βx (x(t))R−1 βxT (x(t))Ŵ (t) (34)

Ṡ(t) = −S(t)βx (x(t))R−1 βx (x(t))T S(t) + Q

where kkf is the positive learning rate of this KF based version of ALR.

We compare the performance of the adaptive law in Eq. (33) with the adaptive

law in Eq. (34), using the same wing rock dynamics as presented in the previous

30
subsection, with σ = 0.001. Figures 11-13 show the standard ALR adaptive control
responses using Eq. (33) with several different gains. It is clear from these figures

that the system response is reasonable only for the case γ = 100, kgrad = 100 as

depicted in Figure 13. Figures 14-16 show the KF based ALR (KF-ALR) adaptive

control responses using Eq. (34) for the same gain settings. It is obvious from these

figures that for all cases in which kkf > 0, KF-ALR is able to suppress the uncertainty

successfully, even with the lowest setting for γ.

8 30
Eq. (33), γ=10, kgrad=0
6 20
Eq. (33), γ=10, k =10
grad
x2 [degree/s]

4 10
x1 [degree]

2 0

0 −10

−2 −20

−4 −30
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

4 1.4

3 1.2
2
1
Control Signal

1
||W βx||

0.8
0
T

0.6
−1
0.4
−2

−3 0.2

−4 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 11: Standard ALR based adaptive control response with uncertainty using
Eq. (33) with γ = 10 and kgrad = 10.

31
 8 10
Eq. (33), γ=100, k =0
grad
6 Eq. (33), γ=100, k =10
grad 5

x2 [degree/s]
x1 [degree]
4
0
2

−5
0

−2 −10
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

4 1.4

3 1.2
2
1
Control Signal

||W βx||
0.8
0

T
0.6
−1
0.4
−2

−3 0.2

−4 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 12: Standard ALR based adaptive control response with uncertainty using
Eq. (33) with γ = 100 and kgrad = 10.

8 8
Eq. (33), γ=100, k =0
grad 6
6 Eq. (33), γ=100, kgrad=100
4
x2 [degree/s]
x1 [degree]

4 2

0
2 −2

−4
0
−6

−2 −8
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 1.4

2 1.2

1
Control Signal

1
||W βx||

0.8
0
T

0.6
−1
0.4

−2 0.2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 13: Standard ALR based adaptive control response with uncertainty using
Eq. (33) with γ = 100 and kgrad = 100.

32
 8 30
Eq. (34), γ=10, k =0
kf
6 20
Eq. (34), γ=10, k =10
kf

x2 [degree/s]
4 10

x1 [degree]
2 0

0 −10

−2 −20

−4 −30
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 1.4

2 1.2

1
Control Signal

||W βx||
0.8
0

T
0.6
−1
0.4

−2 0.2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 14: KF-ALR based adaptive control response with uncertainty using Eq. (34)
with γ = 10 and kgrad = 10.

8 8
Eq. (34), γ=100, k =0
kf 6
6 Eq. (34), γ=100, kkf=10
4
x2 [degree/s]
x1 [degree]

4 2

0
2 −2

−4
0
−6

−2 −8
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 1.4

2 1.2

1
Control Signal

1
||W βx||

0.8
0
T

0.6
−1
0.4

−2 0.2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 15: KF-ALR based adaptive control response with uncertainty using Eq. (34)
with γ = 100 and kgrad = 10.

33
 8 8
Eq. (34), γ=100, k =0
kf 6
6 Eq. (34), γ=100, k =100
kf 4

x2 [degree/s]
x1 [degree]

4 2

0
2 −2

−4
0
−6

−2 −8
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

3 1.4

2 1.2

1
Control Signal

1
||W βx||

0.8
0
T

0.6
−1
0.4

−2 0.2

−3 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 16: KF-ALR based adaptive control response with uncertainty using Eq. (34)
with γ = 100 and kgrad = 100.

34
2.4.3 A New ue –Modification for Input Constraints

A new treatment of actuator dynamics and limits is proposed using the KF based

optimization procedure. We present the following results for illustration purposes

without proof of stability. For this purpose, let u(t) be the control commanded to the

actuator of a nonlinear uncertain system, and let us (t) be the output of the actuator.

Defining ue (t) = us (t) − u(t), our aim is to approximately apply the following linear

constraint

1
|ue (t)| 2 Ŵ (t) = 0 (35)

to the adaptive law with σ–modification, given by

˙
Ŵ (t) = γ[β(x(t))eT (t)P B − σ Ŵ (t)], (36)

Applying the form in (35) to (21) and (22), results in:

˙
Ŵ (t) = −|ue (t)|S(t)R−1 Ŵ (t)

Ṡ(t) = −|ue (t)|S(t)R−1 S(t) + Q

Remark 2.7. Since Q, R, and S0 are chosen in the form q̄I, r̄I, and s̄o I, respectively,

then the above equations reduce to:

˙ |ue (t)|s(t)
Ŵ (t) = − Ŵ (t) (37)
r̄
s2 (t)
ṡ(t) = −|ue (t)| + q̄ (38)
r̄

The adaptive law in (36) augmented with the ue –modification term in (37) then

becomes
h |ue (t)|s(t) i
˙
Ŵ (t) = γ β(x(t))eT (t)P B − σ Ŵ (t) − σue Ŵ (t) (39)
r̄

where σue is a positive learning rate. The main assumption is that the actuator output

is known or can be estimated to sufficient accuracy using a model for the actuator.

35
 We compare ue –modification with the method of hedging [39, 40], that uses the
adaptive law in (36) and modifies the reference model in (4) with a so-called hedge

signal. The same wing rock adaptive control problem is used, with the design param-

eters γ = 1, σ = 3, and σue = 50. Furthermore, the following actuator dynamics are

included

u̇s (t) = −τ us (t) + τ u(t) (40)

where τ = 1/3.

Figure 17 shows that the presence of actuator dynamics significantly degrades per-

formance when using the adaptive law in Eq. (36). In Figures 18 and 19, hedging was
used to modify the reference model dynamics. Figure 13 shows that the response with

hedging becomes less oscillatory. However, Figure 18 shows that hedging modifies the

reference model into a response that is more oscillatory. Figure 20 shows the response

when ue –modification is employed. The response is improved without modifying the

reference model. Figure 21 compares the results of hedging and ue –modification for

the case where, in addition, an amplitude limit of ±0.25 is applied. Both methods
improve the system performance significantly when subjected to a hard limit, however

ue –modification achieves this without changing the reference model. Finally, Figure

22 shows the performance of the adaptive controller with ue –modification, for the

case where the actuator dynamics include an input time delay of 0.1 seconds. The

system response is unstable without ue –modification, and this response is significantly

improved when ue –modification term is employed.

36
 35 15
Eq.(36), γ=1, σ=3, without actuator
30 Eq.(36), γ=1, σ=3, with actuator 10
25

x2 [degree/s]
5

x1 [degree]
20

15 0

10
−5
5
−10
0

−5 −15
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

0.2 0.6

0.1 0.4
Control Amplitude

0 0.2

Control Rate
−0.1 0

−0.2 −0.2

−0.3 −0.4

−0.4 −0.6

−0.5 −0.8
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

Figure 17: Adaptive control response without and with actuator dynamics.

40 15
Eq.(36), γ=1, σ=3
Eq.(36), γ=1, σ=3, with hedging 10
30
x2 [degree/s]

5
x1 [degree]

20
0
10
−5

0
−10

−10 −15
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

0.3 0.6

0.2 0.4
Control Amplitude

0.1
0.2
Control Rate

0
0
−0.1
−0.2
−0.2
−0.4
−0.3

−0.4 −0.6

−0.5 −0.8
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

Figure 18: Performance of adaptive controller with hedging, with actuator dynamics.

37
 30
Eq.(36), γ=1, σ=3
25 Eq.(36), γ=1, σ=3, with hedging

20

xm [degree]
15

10
1

−5
0 5 10 15 20 25
Time [s]

10

5
xm [degree/s]

0
2

−5

−10
0 5 10 15 20 25
Time [s]

Figure 19: Response of reference model without and with hedging, with actuator
dynamics.

35 15
Eq.(36), γ=1, σ=3
30 Eq.(39), γ=1, σ=3, σu =50 10
e
25
x2 [degree/s]

5
x1 [degree]

20

15 0

10
−5
5
−10
0

−5 −15
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

0.2 0.6

0.1 0.4
Control Amplitude

0 0.2
Control Rate

−0.1 0

−0.2 −0.2

−0.3 −0.4

−0.4 −0.6

−0.5 −0.8
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

Figure 20: Performance of adaptive controller with ue -modification, with actuator

dynamics.

38
 50 20
Eq.(36), γ=1, σ=3
40 Eq.(36), γ=1, σ=3, with hedging
10
30 Eq.(39), γ=1, σ=3, σu =50
e

x2 [degree/s]
x1 [degree]
20 0
10

0 −10

−10
−20
−20

−30 −30
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

0.3 0.8

0.2 0.6
Control Amplitude

0.4
0.1

Control Rate
0.2
0
0
−0.1
−0.2
−0.2
−0.4
−0.3 −0.6

−0.4 −0.8
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

Figure 21: Comparison of adaptive control responses with hedging versus with ue -
modification, with actuator dynamics and limits.

40 60

40
20
20
x2 [degree/s]
x1 [degree]

0
0

−20
−20
Eq.(36), γ=1, σ=3
−40
Eq.(39), γ=1, σ=3, σu =50
−40 e

Eq.(39), γ=1, σ=3, σu =250 −60

e

−60 −80
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

2 4

3
1
Control Amplitude

2
Control Rate

0 1

0
−1 −1

−2
−2
−3

−3 −4
0 5 10 15 20 25 0 5 10 15 20 25
Time [s] Time [s]

Figure 22: Performance of adaptive controller with ue -modification, with actuator

dynamics and time delay.

39
 Figure 23: TwinSTAR flight test vehicle.

2.5 Conclusion

The intent of this chapter has been to present a procedure that approximately enforces

linear constraints to modify an existing adaptive control design by using a Kalman

filter optimization approach. The form of the constraint is chosen so that Kalman

filter versions of existing modification terms are obtained. The simulation results on

a model of wing rock dynamics illustrate the presented theory and show significant

improvement over other gradient based modification approaches. One key difference

in the Kalman filter based approach to modification is that the resulting gain on the

modification term is time varying. The proposed approach to deriving a modification

term for an adaptive controller can be used in place of all modification terms that

can be equivalently viewed as the gradient of a norm measure of a linear constraint

on the adaptation gains. Finally, the proposed approach has been flight tested on

TwinSTAR, a conventional twin engine RC aircraft shown in Figure 23, and the

results are reported in Ref. 14.

40
 CHAPTER III

Kalman Filter Based Adaptive Control

3.1 Introduction

This chapter presents an approach to construct a novel weight adaptation law us-

ing Kalman filter optimization. This approach is an extension of the Kalman filter
modification approach presented in the previous chapter. Instead of optimizing the

modification gain, the adaptation gain is optimized. This approach results in an opti-

mization based, time-varying adaptation gain that achieves better performance than

adaptive laws that employ a fixed gain.

3.2 Preliminaries

In this section, we consider the problem of characterizing adaptive control laws for

a class of uncertain systems given by (1), where the full state is available for feed-
back, the control input u(t) is restricted to the class of admissible controls consisting

of measurable functions, and Assumption 2.3 holds. In order to achieve trajectory

tracking, we construct a reference system given by (4), where Assumption 2.2 holds.

We consider the control law in (5) with the nominal control law given by (2) and the

adaptive control law given by (6) where an estimate of W obtained from the adaptive

weight update law

˙  ˙ 
Ŵ (t) = γ β(x(t))eT (t)P B + Ŵm (t) , (41)

41
where γ is a positive fixed gain, e(t) ≡ x(t) − xm (t) is the error, P ∈ IRn×n is a
˙
positive-definite solution of the Lyapunov equation given by (8), and Ŵm (t) ∈ IRs×m
˙
is a modification term, for example Ŵm (t) = −σ Ŵ (t) for σ–modification term [37]
˙
and Ŵm (t) = −σ|e(t)|Ŵ (t) for e–modification term [60], where σ is a positive fixed

gain.

We stated the UUB property of the closed loop system error defined by e(t) ≡

x(t) − xm (t) and W̃ (t) ≡ W − Ŵ (t) in Theorem 2.1 for (41) when e–modification is
employed. For completeness, we now state and prove the UUB property for (41) when

σ–modification is employed. For this purpose, the weight update error dynamics can

be expressed as:

˙ (t) = −γ β(x(t))eT (t)P B − σ Ŵ (t)

W̃ (42)

Let ζ(t), Br , Ωα , and Dx as in Theorem 2.1 (see Figure 2). Define P̃ = diag[P, γ −1 I],
p p
Θe1 = ||W ||2 + ||P B||2ǫ2 /(2ξσ), ΘW̃1 = (2ξσ||W ||2 + ||P B||2ǫ2 )/(λmin(Qm ) − 2ξ),
1 2
λmax (P )Θ2e1 + Θ
γ W̃1
and r2σ = λmin (P̃ )
, where λmin (Qm ) > 2ξ, ξ > 0.

Theorem 3.1. Consider the uncertain system given by (1) subject to Assumption
2.3. Consider, in addition, the feedback control law given by (5), with the nominal

feedback control component given by (2) subject to Assumption 2.2, and with the

adaptive feedback control component given by (6) that has a weight update law in
˙
the form (41) with Ŵm (t) = −σ Ŵ (t). If ζ(t0 ) ∈ Ωα and rα > r σ , then ζ(t) is UUB

with ultimate bound r σ .

Proof. Consider the Lyapunov-like function candidate given by (13) where P > 0
satisfies (8). The time derivative of (13) along the trajectories of (11) and (42) can

42
be expressed as

V̇(·) = eT (t)P [Am e(t) + B W̃ T (t)β(x(t)) + Bε(x(t))] − tr[W̃ T (t)β(x(t))eT (t)P B]

+σtr[W̃ T (t)Ŵ (t)]

1 T
= e (t)[AT T T
m P + P Am ]e(t) + σtr[W̃ (t)Ŵ (t)] + e (t)P Bε(x(t))
2
1
= − eT (t)Qm e(t) + σtr[W̃ T (t)(W − W̃ (t))] + eT (t)P Bε(x(t))
2
≤ −c1 |e(t)|2 + c2 |e(t)| − c3 ||W̃ (t)||2 + c4
c22
≤ −(c1 − ξ)|e(t)|2 − c3 ||W̃ (t)||2 + c4 + , (43)
4ξ
λmin (Qm ) σ σ||W ||2
where c1 = 2
> ξ > 0, c2 = ||P B||ǫ > 0, c3 = 2
> 0, and c4 = 2
> 0.
Consequently, either |e(t)| ≥ Θe1 or ||W̃ (t)|| ≥ ΘW̃1 renders V̇(e(t), W̃ (t)) < 0, where
r r
c2 c2
c4 + 42 c4 + 42
Θe 1 = c1 −1
and ΘW̃1 = c3
. It follows now identical to the proof of Theorem

2.1 that e(t) and W̃ (t) are UUB with ultimate bound r σ . 

3.3 Motivation

For motivational purposes, we show that σ– and e–modification based gradient weight

adaptation laws can be obtained by enforcing the constraint

Ŵ T (t)Φ1 (t, x(t), e(t)) = Φ2 (t, x(t), e(t)) (44)

on the weight estimates, where Φ1 (·) : [0, ∞) × IRn × IRn → IRs×l and Φ2 (·) : [0, ∞) ×

IRn × IRn → IRm×l .

Remark 3.1. The form of the constraint in (44) differs than the form of the

constraint in (15). If Φ2 (·) = 0 in (44), then (44) reduces to (15).

Lemma 3.1. The σ–modification based gradient adaptive control law given by
˙
(41) with Ŵm (t) = −σ Ŵ (t) can be obtained by using a gradient approach to enforcing

43
the constraint given by (44) on the weight estimates, where:
√
Φ1 (·) ≡ σI (45)
√
Φ2 (·) ≡ β(x(t))eT (t)P B/ σ, σ>0 (46)

Proof. Consider the cost

1
J (Ŵ (t)) = ||Ŵ T (t)Φ1 (·) − Φ2 (·)(x(t))||2 (47)
2
where the negative gradient of J (Ŵ (t)) with respect to Ŵ (t) gives
∂J (Ŵ (t))
− = β(x(t))eT (t)P B − σ Ŵ (t) (48)
∂ Ŵ (t)
where (45) and (46) are used. Multiplying both sides of (48) by γ results in σ–
˙
modification based adaptive control law in (41) with Ŵm (t) = −σ Ŵ (t). 

Lemma 3.2. The e–modification based gradient adaptive control law given by
˙
(41) with Ŵm (t) = −σ|e(t)|Ŵ (t) can be obtained by using a gradient approach to
enforcing the constraint given by (44) on the weight estimates, where:
p
Φ1 (·) ≡ σ|e(t)|I (49)
p
Φ2 (·) ≡ β(x(t))eT (t)P B/ σ|e(t)|, σ>0 (50)

Proof. Consider the cost given by (47) where the negative gradient of J (Ŵ (t))

with respect to Ŵ (t) gives

∂J (Ŵ (t))
− = β(x(t))eT (t)P B − σ|e(t)|Ŵ (t) (51)
∂ Ŵ (t)
where (49) and (50) are used. Multiplying both sides of (51) by γ results in e–
˙
modification based adaptive control law in (41) with Ŵm (t) = −σ|e(t)|Ŵ (t). 

It is shown in Lemmas 3.1 and 3.2 that σ– and e–modification based gradient

adaptive control laws can be obtained by imposing the constraint given by (44) on
the weight estimates subject to (45) and (46), and (49) and (50), respectively. In the

next section, we introduce a different treatment for a general class of linear constraints

to obtain Kalman filter based adaptive control (KF-AC) laws.

44
3.4 Kalman Filter Based Adaptive Control Formulation

The KF-AC architecture is applicable to most direct adaptive control schemes avail-

able in the literature in full-state feedback control or output feedback control forms.
For simplicity and without loss of generality, we consider the full-state feedback for-

mulation. As highlighted in Lemmas 3.1 and 3.2, (41) can be viewed as a gradient

approach for enforcing the linear constraint (44) on the weight estimates.

Next, the problem of estimating W , so that it satisfies the the linear constraint in

(44), will be formulated as an optimization problem. For this purpose, (44) can be

expressed in the following equivalent vector form

Φ̃T T
1 (t, x(t), e(t))ω(t) = Φ̃2 (t, x(t), e(t)) (52)

where Φ̃1 (·) ∈ IRms×ml , Φ̃2 (·) ∈ IR1×ml , and ω(t) = vec(Ŵ (t)) ∈ IRms .

Define the stochastic process

ω̇(t) = q(t), (53)

z(t) = Φ̃T
1 (·)ω(t) + r(t), (54)

where q(t) and r(t) are zero-mean, Gaussian, white noise processes with covariances


E q(t)q T (τ ) = Q̄δ(t − τ ), Q̄ ∈ IRms×ms > 0 (55)

E r(t)r T (τ ) = R̄δ(t − τ ), R̄ ∈ IRml×ml > 0 (56)

and z(t) is regarded as a measurement. The estimate of z(t) is

ẑ(t) = Φ̃T
1 (·)ω̂(t) (57)

where ω̂(t) is an estimate of ω(t).

The Kalman filter associated with this problem formulation is given by [19]:

 
˙
ω̂(t) = S̄(t)Φ̃1 (·)R̄−1 z(t) − ẑ(t) , ω̂(0) = 0 (58)
˙
S̄(t) = −S̄(t)Φ̃1 (·)R̄−1 Φ̃T
1 (·)S̄(t) + Q̄, S̄(0) = S̄0 > 0 (59)

45
where S̄(t) ∈ IRms×ms . Since the objective is to approximately satisfy the constraint
(44), the logical choice for z(t) when employing this estimator is z(t) = Φ̃T
2 (·). Fur-

thermore, choosing R̄ = Im×m ⊗ R, with R ∈ IRl×l > 0, and Q̄ = Im×m ⊗ Q, with

Q ∈ IRs×s > 0, Lemma 2.1 of Section 2.3 shows that (58) and (59), with z(t) = Φ̃T
2 (·),

can equivalently be expressed as:

˙  
Ŵ (t) = −S(t)Φ1 (·)R−1 ΦT T
1 (·)Ŵ (t) − Φ2 (·) , Ŵ (0) = 0 (60)

Ṡ(t) = −S(t)Φ1 (·)R−1 ΦT

1 (·)S(t) + Q, S(0) = S0 > 0 (61)

where S(t) ∈ IRs×s .

Remark 3.2. The Riccati equation in (61) can be associated with the following

system of equations:

ẇi = q1 (t)

zi (t) = ΦT
1 (·)wi + r1 (t)

in which wi are the process dynamics associated with the i-th column of W , zi is the
 
corresponding measurement vector, E q1 (t)q1T (τ ) = Qδ(t − τ ), and E r1 (t)r1T (τ ) =
Rδ(t − τ ).

Assumption 3.1. The observability gramian

Z t
W0 (t0 , t) = Φ1 (t, x(t), e(t))ΦT
1 (t, x(t), e(t))dt (62)
t0

of the system defined in Remark 3.2 satisfies the following condition for uniform
complete observability [43]:

0 < α0 (σ)Il ≤ W0 (t0 , t0 + σ) ≤ α1 (σ)Il (63)

for all t0 , and for some fixed constant σ > 0.

Remark 3.3. The condition in Assumption 3.1 is sufficient to show that S(t) is

positive-definite, and lower and upper bounded, independent of t0 , if Φ1 (t, x(t), e(t))

is bounded pointwise in time.

46
3.5 Stability Analysis

The stability analysis of the KF-AC law given by (60) and (61) depends on the

constraint given by (44), and hence, it is problem dependent. Therefore, we need to

limit our presentation, without loss of generality, by giving the KF-AC forms of σ–

and e–modification based adaptive control laws.

From the constraint (44) with (45) and (46), and the equations (60) and (61),

σ–modification based KF-AC (KFσ –AC) law can be given by:

˙  
Ŵ (t) = S(t)R−1 β(x(t))eT (t)P B − σ Ŵ (t) (64)

Ṡ(t) = −σS(t)R−1 S(t) + Q (65)

Similarly, from the constraint (44) with (49) and (50), and the equations (60) and
(61), e–modification based KF-AC (KFe –AC) law can be given by:

˙  
Ŵ (t) = S(t)R−1 β(x(t))eT (t)P B − σ|e(t)|Ŵ (t) (66)

Ṡ(t) = −σ|e(t)|S(t)R−1 S(t) + Q (67)

Remark 3.4. For the KFσ –AC law, the observability gramian given by (62)
Rt
becomes W0 (t0 , t) = t0 σdτ Is , and hence, Assumption 3.1 is satisfied. Furthermore,

for the KFe –AC law, the observability gramian given by (62) becomes W0 (t0 , t) =
Rt
t0
σ|e(τ )|dτ Is , and it is positive-definite for all t ≥ t0 as long as |e(τ )| =
6 0 over the
interval [t0 , t]. It will be necessary to prove that |e(t)| remains bounded pointwise in

time using arguments similar to those given in the proof of Theorem 2.2.

Remark 3.5. S(t)R−1 is a time-varying and bounded Kalman filter gain in the

KFσ –AC and KFe –AC laws. However, for the following results, we require it to be

positive-definite and symmetric as well. For this purpose, we set R = rI, r > 0.

The weight update error dynamics for KFσ –AC law can be given using (64) as:

˙ (t) = −S(t)R−1 β(x(t))eT (t)P B − σ Ŵ (t)

W̃ (68)

47
Define ζ(t), Br , Ωα , and Dx as in Theorem 2.1 (see Figure 2). Further define
p p
P̃KF (t) = diag[P, Γ−1 (t)], Γ(t) = S(t)R−1 = S(t)/r, Θe2 = v3 /v1 , ΘW̃2 = v3 /v2 ,
λmax (P )Θ2e2 + Γ̄−1 Θ2
max
and r2σKF = λmin (P̃ KF )
W̃2
, where P̃ KF = diag[P, Γ̄−1 ], Γ̄ = S(t)
{Γ(t)}, v1 =
λmin (Qm )
2
− ξ > 0, ξ > 0, v2 = 12 λmin (QΓ̄−1 ), and v3 = σ2 ||W ||2 + 4ξ1 ||P B||2ǫ2 . The next

theorem states the UUB property of KFσ –AC law.

Theorem 3.2. Consider the uncertain system given by (1) subject to Assumption

2.3. Consider, in addition, the feedback control law given by (5), with the nominal
feedback control component given by (2) subject to Assumption 2.2, and with the

adaptive feedback control component given by (6) together with KFσ –AC law given

by (64). If ζ(t0 ) ∈ Ωα and rα > r σKF , then ζ(t) is UUB with ultimate bound r σKF .

Proof. Consider the Lyapunov function candidate

1 1
V(e(t), W̃ (t)) = eT P e + tr[W̃ (t)T Γ−1 (t)W̃ (t)] (69)
2 2

where P > 0 satisfies (8). The time derivative of (69) along the trajectories of (11)
and (68) can be expressed as

1
V̇(·) = − eT (t)Qm e(t) + eT (t)P Bε(x(t)) + σtr[W̃ T (t)(W − W̃ (t))]
2
1
− tr[W̃ T (t)Γ−1 (t)Γ̇(t)Γ−1 (t)W̃ (t)]
2
1  σ σ 1
≤ − λmin (Qm ) − ξ |e(t)|2 − ||W̃ (t)||2 + ||W ||2 + ||P B||2ǫ2
2 2 2 4ξ
1
− tr[W̃ T (t)Γ−1 (t)Γ̇(t)Γ−1 (t)W̃ (t)] (70)
2

where ξ > 0. Using Γ−1 (t)Γ̇(t)Γ−1 (t) = rQS −1 (t) − σIs in (70) gives:

V̇(·) ≤ −v1 |e(t)|2 − v2 ||W̃ (t)||2 + v3 (71)

Consequently, either |e(t)| ≥ Θe2 or ||W̃ (t)|| ≥ ΘW̃2 renders V̇(·) < 0. Therefore,

e(t) and W̃ (t) are uniformly bounded. Define Ωβ = {ζ(t) ∈ Br : ζ(t)T P̃KF (t)ζ(t) ≤

λmax (P )Θ2e2 + Γ−1 (t)Θ2W̃ (t)}. Then, Ωα is a positively invariant set if Ωβ ⊂ Ωα . This
2

requires that λmax (P )Θ2e2 + Γ−1 (t)Θ2W̃ < α. The minimum size of Br that ensures

48
this condition is rα > r σKF . Therefore, if ζ(t0) ∈ Ωα , then ζ(t) is UUB with ultimate
bound rσKF . 

We next state the UUB property of the KFe –AC law. For this purpose, the weight

update error dynamics for KFe –AC law can be given using (66) as:

˙ (t) = −S(t)R−1 β(x(t))eT (t)P B − σ|e(t)|Ŵ (t)

W̃ (72)

p λmax (P )Θ2e3 + Γ̄−1 Θ2

Define Θe3 = v̄3 /v̄1 , ΘW̃3 = v̄3 /v̄2 , and r 2eKF = λmin (P̃ KF )
W̃3
, where v̄1 =
λmin (Qm )
2
> 0, v̄2 (t) = 12 λmin (QΓ̄−1 ), and v̄3 = 21 σ||W ||2 + ||P B||ǫ.

Theorem 3.3. Consider the uncertain system given by (1) subject to Assumption

2.3. Consider, in addition, the feedback control law given by (5), with the nominal
feedback control component given by (2) subject to Assumption 2.2, and with the

adaptive feedback control component given by (6) together with KFe –AC law given

by (66). If ζ(t0 ) ∈ Ωα and rα > reKF , then either ζ(t) is UUB with ultimate bound

r eKF , or e(t) = 0 for all t ≥ T and W̃ (t) is bounded.

Proof. Consider the Lyapunov function candidate (69) where P > 0 satisfies

(8). Applying similar arguments as in the proofs of Theorems 2.1 and 3.2, the time
derivative of (69) along the trajectories of (11) and (72) can be expressed as:

V̇(·) ≤ −|e(t)|(v̄1 |e(t)| + v̄2 ||W̃ (t)||2 − v̄3 ) (73)

Note that V̇(·) is negative as long as the term in braces is positive and e(t) 6= 0. If
˙ (t) = 0 from (72). It now follows
e(t) = 0 for all t ≥ T , then W̃ (t) is bounded since W̃

identical to the proofs of Theorems 2.2 and 3.2 that if ζ(t0) ∈ Ωα , then ζ(t) is UUB

with ultimate bound r eKF . 

3.6 Examples on a Model of Wing Rock Dynamics

In this section we apply our results to a model of wing rock dynamics given by (32)

with α1 = 0.9814, α2 = 1.5848, α3 = −0.6245, α4 = 0.0095, and α5 = 0.0215.

49
 30
x1(t)
20
x2(t)

States
10

−10
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
Control Signal 0.1
u(t)
0

−0.1

−0.2

−0.3
0 20 40 60 80 100 120 140 160 180 200
Time (sec)

Figure 24: Nominal control response.

The control objective is to minimize the oscillations of the wing rock dynamics in

order to stabilize the system at the zero trim condition. We selected the initial state

values as x(0) = [15o , 5o /s]T . The reference system is selected to be second order

with a natural frequency of 1.0 rad/s, and a damping of 0.8, and to have unity low

frequency gain from r(t) to x1 (t). This corresponds to choosing K1 = [1, 1.6] and
K2 = 1. The form in Assumption 2.3 was used to approximate the uncertainty with

β(x(t)) = [β1 (x(t)), β2 (x(t))]T , where βi (x) = 1/(1 + e−xi ), i = 1, 2, are sigmoidal

type basis functions. Furthermore, we chose Qm = I2 for (8) and Q = I2 , R = 0.01I2 ,

S0 = I2 for the KFe –AC law given by (66) and (67).

Fig. 24 shows the nominal control response with the matched uncertainty Π(x(t)),

where uad (t) = 0. Figs. 25–28 presents the standard, gradient, e–modification based
adaptive controller results on this system for different gains. It is obvious from all

these figures that the achieved system performance is not desirable due to high fre-

quency oscillations in the system responses. Fig. 29 shows the performance of KFe –

AC law on this system, where its time-varying Kalman filter gain is also given in Fig.

30. The performance is significantly improved relative to the results obtained using

the standard, gradient, e–modification term. Hence, this approach leads to smaller
tracking errors without incurring high-frequency oscillations in the system response.

50
 20 x1(t)
15
x2(t)

States
10
5
0

0 5 10 15 20 25 30 35 40
Time (sec)
0.1
u(t)
Control Signal

−0.1

−0.2

−0.3
0 5 10 15 20 25 30 35 40
Time (sec)
0.06
Weight Estimate

|| W(t) ||
0.04

0.02

0
0 5 10 15 20 25 30 35 40
Time (sec)

Figure 25: Standard e- modification based adaptive controller, γ = 1 and γ ×σ = 40.

20
x1(t)
10
x2(t)
States

−10

−20
0 5 10 15 20 25 30 35 40
Time (sec)
1
u(t)
Control Signal

0.5

−0.5

−1
0 5 10 15 20 25 30 35 40
Time (sec)
1.5
Weight Estimate

|| W(t) ||
1

0.5

0
0 5 10 15 20 25 30 35 40
Time (sec)

Figure 26: Standard e- modification based adaptive controller, γ = 10 and γ×σ = 40.

51
 20
x1(t)
10
x2(t)

States
0

−10

−20
0 5 10 15 20 25 30 35 40
Time (sec)
2
u(t)
Control Signal

−1

−2
0 5 10 15 20 25 30 35 40
Time (sec)
6
Weight Estimate

|| W(t) ||
4

0
0 5 10 15 20 25 30 35 40
Time (sec)

Figure 27: Standard e- modification based adaptive controller, γ = 100 and γ × σ =

40.

20
x1(t)
10 x2(t)
States

−10
0 5 10 15 20 25 30 35 40
Time (sec)
2
u(t)
Control Signal

−1

−2
0 5 10 15 20 25 30 35 40
Time (sec)
8
Weight Estimate

|| W(t) ||
6

0
0 5 10 15 20 25 30 35 40
Time (sec)

Figure 28: Standard e- modification based adaptive controller, γ = 1000 and γ ×σ =

40.

52
 20
x1(t)
15
x2(t)
States

10

0 5 10 15 20 25 30 35 40
Time (sec)
0.1
u(t)
Control Signal

−0.1

−0.2

−0.3
0 5 10 15 20 25 30 35 40
Time (sec)
0.06
Weight Estimate

|| W(t) ||
0.04

0.02

0
0 5 10 15 20 25 30 35 40
Time (sec)

Figure 29: KFe –AC law, σ = 40.

1000
|| S(t) R−1 ||
H2 Gain

500

0
0 10 20 30 40 50 60 70 80
Time (sec)

Figure 30: Adaptation gain Γ(t) for the KFe –AC law in Figure 29.

53
3.7 Conclusion

The intent of this chapter has been to present a Kalman filter based approach to
deriving a novel weight adaptation law. A model of wing rock dynamics illustrates

the presented theory. The new controllers showed significant improvement over their

gradient based counterparts. One key difference in the Kalman filter based approach

is that the resulting adaptation gain is time varying, and requires minimal tuning of

the design parameters. The proposed approach can be used in place of all adaptive

control laws that can be equivalently viewed as the gradient of a norm of the error
in the linear constraint. Finally, the proposed approach has been flight tested on

TwinSTAR and the results are reported in Ref. 15.

54
 CHAPTER IV

Derivative-Free Adaptive Control: The Full-State Feedback

Case

4.1 Introduction

A derivative-free, delayed weight update law is developed for model reference adap-

tive control of continuous-time uncertain systems, without assuming the existence

of constant ideal weights and without requiring discretization. Using a Lyapunov-

Krasovskii functional it is proven that the error dynamics are uniformly ultimately

bounded, without the need for modification terms in the adaptive law. Estimates for

the ultimate bound and the exponential rate of convergence to the ultimate bound are

provided. We also discuss employing various modification terms for further improving

performance and robustness of the adaptively controlled system. Examples illustrate

that the proposed derivative-free model reference adaptive control (DF-MRAC) law

is advantageous for applications to systems that can undergo a sudden change in

dynamics.

4.2 Preliminaries

In this chapter, we consider the uncertain system given by (1), in which ∆(x(t)) is

replaced by ∆(t, x(t)), the full state is available for feedback, and the control input

u(t) is restricted to the class of admissible controls consisting of measurable functions.

In order to achieve trajectory tracking, we construct a reference system given by (4),

where Assumption 2.2 holds. We consider the control law in (5) with the nominal

55
control law given by (2) and the adaptive control law given by (6).

4.3 Derivative-Free Adaptive Control

The following assumption strengthens Assumption 2.3 by setting ε(x) = 0, which can

be justified under the assumption that time-variation is allowed in the unknown ideal

weight matrix.

Assumption 4.1. The matched uncertainty in (1) can be linearly parameterized

as

∆(t, x) = W T (t)β(x), x ∈ Dx (74)

where W (t) ∈ IRs×m is an unknown time-varying weight matrix that satisfies kW (t)k

≤ w ∗ and β : IRn → IRs is a vector of known functions of the form β(x) = [β1 (x),
β2 (x), . . . , βs (x)]T ∈ IRs .

Remark 4.1. Assumption 4.1 expands the class of uncertainties that can be

represented by a given set of basis functions. That is, an adaptive law designed subject

to Assumption 4.1 can be more effective than an adaptive law designed subject to

Assumption 2.3 in dominating a wider class of uncertainties, due to the fact that

time-variation is allowed in the unknown ideal weight matrix.

Remark 4.2. Assumption 4.1 does not place any restriction on the time derivative

of the weight matrix. However the degree of time dependence will depend on how

β(x) is chosen.

The state error dynamics in (11) with ε(x) = 0 is now given by

ė(t) = Am e(t) + B W̃ T (t)β(x(t)) (75)

where W̃ (t) ≡ W (t)−Ŵ (t) is the weight update error. The following theorem presents

the main result of this section.

56
 Theorem 4.1. Consider the uncertain system given by (1) subject to Assumption
4.1. Consider, in addition, the feedback control law given by (5), with the nominal

feedback control component given by (2) subject to Assumption 2.2, and with the

adaptive feedback control component given by (6) that has a derivative-free weight

update law in the form

Ŵ (t) = Ω1 Ŵ (t − τ ) + Ω̂2 (t) (76)

where τ > 0, and Ω1 ∈ IRs×s and Ω̂2 : IRn × IRn → IRs×m satisfy:

0 ≤ ΩT
1 Ω1 < I (77)

Ω̂2 (t) = κ2 β(x(t))eT (t)P B, κ2 > 0 (78)

with P ∈ IRn×n satisfying (8) for any symmetric Qm > 0. Then, e(t) and W̃ (t) are
UUB.

Proof. Using (76) and defining

Ω2 (t) ≡ W (t) − Ω1 W (t − τ ) (79)

where ||Ω2 (t)|| ≤ δ ∗ , δ ∗ = w ∗(1 + ||Ω1 (t)||), the weight update error can be rewritten

as:

W̃ (t) = Ω1 W̃ (t − τ ) + Ω2 (t) − Ω̂2 (t) (80)

The state tracking error dynamics in (75) can be rewritten using (80) as:

 T
ė(t) = Am e(t) + B Ω1 W̃ (t − τ ) − Ω̂2 (t) + Ω2 (t) β(x(t)) (81)

To show that the closed-loop system given by (80) and (81) is UUB, consider the

Lyapunov-Krasovskii functional [31]

Z t
T
 
V(e(t), W̃t ) = e (t)P e(t) + ρ tr W̃ T (s)W̃ (s)ds (82)
t−τ

57
where ρ > 0 and W̃t represents W̃ (t) over the time interval t − τ to t. The directional
derivative of (82) along the closed-loop system trajectories of (80) and (81) is given

by

V̇(e(t), W̃t ) = −eT (t)Qm e(t) + 2eT (t)P B[Ω1 W̃ (t − τ )]T β(x(t))

−2eT (t)P B Ω̂T T T

2 (t)β(x(t)) + 2e (t)P BΩ2 (t)β(x(t))

+ρ tr −ξ W̃ T (t)W̃ (t) + η W̃ T (t)W̃ (t)

−W̃ T (t − τ )W̃ (t − τ ) (83)

where η = 1 + ξ. In what follows we impose the restriction ξ ≥ 0.

Using (80) to expand the term tr[η W̃ T (t)W̃ (t)] in (83) produces

V̇(e(t), W̃t ) = −eT (t)Qm e(t) + 2eT (t)P B[Ω1 W̃ (t − τ )]T β(x(t))

−2eT (t)P B Ω̂T T T

2 (t)β(x(t)) + 2e (t)P BΩ2 (t)β(x(t))

+ρ tr −ξ W̃ T (t)W̃ (t) − W̃ T (t − τ )W̃ (t − τ )

+η W̃ T (t − τ )ΩT T T
1 Ω1 W̃ (t − τ ) + η Ω̂2 (t)Ω̂2 (t) + ηΩ2 (t)Ω2 (t)

−2η Ω̂T T T
2 (t)Ω1 W̃ (t − τ ) + 2η W̃ (t − τ )Ω1 Ω2 (t)

−2η Ω̂T
2 (t)Ω2 (t) (84)

Next, consider the fact aT b ≤ γaT a + bT b/4γ, γ > 0, that follows from Young’s

inequality [5, 17] for any vectors a and b. This can be generalized to matrices
   
as tr AT B = vec(A)T vec(B) ≤ γvec(A)T vec(A) + vec(B)T vec(B)/4γ = γtr AT A
 
+tr B T B /4γ, γ > 0, for any matrices A and B with appropriate dimensions. Using

this, we can write

   
tr 2η W̃ T (t − τ )ΩT
1 Ω2 (t) ≤ tr γ W̃ T (t − τ )ΩT 1 Ω1 W̃ (t − τ )
 
+tr η 2 ΩT 2 (t)Ω2 (t)/γ , γ>0 (85)

58
Using (78) with κ2 , 1/ρη > 0 for Ω̂2 (t), and substituting (85) in (84), it follows that

V̇(e(t), W̃t ) ≤ −eT (t)Qm e(t) − κ2 eT (t)P BB T P e(t)β(x(t))T β(x(t))

h   i
−ρξ tr[W̃ T (t)W̃ (t)] − ρ tr W̃ T (t − τ ) I − (η + γ)ΩT Ω
1 1 W̃ (t − τ )
 
+ρ(η + η 2 /γ)tr ΩT 2 (t)Ω2 (t) (86)

Using (77) with κ1 , 1/(η + γ) < 1 for Ω1 yields

V̇(e(t), W̃t ) ≤ −c1 ke(t)k2 − c2 kW̃ (t)k2 − c3 kW̃ (t − τ )k2 + d (87)

where the constants c1 , c2 , c3 , and d are given by :

c1 = λmin (Qm ) > 0 (88)

c2 = ρξ ≥ 0 (89)

c3 = ρλmin (I − κ−1 T
1 Ω1 Ω1 ) > 0 (90)
2
d = ρ(η + η 2 /γ)δ ∗ ≥ 0 (91)

where ρ = 1/κ2 η. If ξ > 0, then since η = κ−1

1 − γ = 1 + ξ, 0 < κ1 < 1, and

γ > 0, it follows that η must lie in the open interval (1, 1/κ1). Either |e(t)| > Ψ1
p
or kW̃ (t)k > Ψ2 or kW̃ (t − τ )k > Ψ3 renders V̇(e(t), W̃t ) < 0, where Ψ1 = d/c1 ,
p p
Ψ2 = d/c2 , and Ψ3 = d/c3 , or equivalently:
q

∗
Ψ1 = δ ρ η + η 2 /γ /λmin (Qm )
q  
∗
= δ / κ2 λmin (Qm ) 1 − ηκ1 (92)
p
Ψ2 = Ψ1 c1 /c2
p
= Ψ1 κ2 λmin (Qm )η/[η − 1] (93)
p
Ψ3 = Ψ1 c1 /c3
q
= Ψ1 κ1 κ2 λmin (Qm )η/λmin(κ1 I − ΩT
1 Ω1 ) (94)

Hence, it follows that e(t) and W̃ (t) are UUB. 

59
 Figure 31: Visualization of the proposed DF-MRAC architecture.

The proposed adaptive control architecture is shown in Figure 31.

Remark 4.3. Using a 1st order Euler method for integration in (7), with τs being

the step size results in:

 
Ŵ (t) = γτs β(x(t))eT (t)P B +Ŵ (t − τs ) (95)

This form of weight update law is identical to the DF-MRAC law in (76), if Ω1 = I,

κ2 = γτs , and τ = τs , with the exception that the choice Ω1 = I is not permitted in

DF-MRAC. In DF-MRAC, Ω1 can be chosen, for example, as ςI where 0 < |ς| < 1,

and τ is not necessarily equal to τs . This added dimension in the tuning process

provides memory to the adaptive law, and can be employed to improve transient

behavior without increasing the effective adaptation gain. As an another example,

using a 1st order Euler method for integration in (41) with the σ-modification term
˙
Ŵm (t) = −σ Ŵ (t) results in:

γτs   1
Ŵ (t) = β(x(t))eT (t)P B + Ŵ (t − τs ) (96)
1 + σγτs 1 + σγτs

60
This form of weight update law is identical to the DF-MRAC law in (76), if Ω1 =
(1 + σγτs )−1 I, κ2 = (1 + σγτs )−1 γτs , and τ = τs . However, τ is not necessarily

equal to τs . The value of τ influences the expressions for the ultimate bound and the

guaranteed exponential rate of convergence to the ultimate bound, as stated below

in Corollaries 4.1 and 4.2.

Remark 4.4. The derivative-free weight update law given by (76) subject to (77)

and (78) does not require a modification term to prove the error dynamics, including
the weight errors, are UUB.
R 0 
Define q(t) ≡ [eT (t), ṽ(t, τ )]T , where ṽ 2 (t, τ ) ≡ tr −τ
W̃ T (s)W̃ (s)ds , and let

Br = {q(t) : ||q(t)|| < rα }, such that Br ⊂ Dq for a sufficiently large compact set Dq .

Then, we have the following corollary.

Corollary 4.1. Under the conditions of Theorem 4.1, an estimate for the ultimate

bound, for the case ξ > 0, is given by

s
λmax (P )Ψ21 + ρτ Ψ22
r= (97)
λmin (P̃ )

where P̃ = diag[P, ρ].

min
Proof. Denote Ωα = {q(t) ∈ Br : q T (t)P̃ q(t) ≤ α}, α = ||q(t)||=r
q T (t)P̃ q(t) =

rα2 λmin (P̃ ). Since

V(e(t), W̃t ) = q T (t)P̃ q(t)

Z t
T
 
= e (t)P e(t) + ρ tr W̃ T (s)W̃ (s)ds (98)
t−τ

it follows that Ωα is an invariant set as long as:

α ≥ λmax (P )Ψ21 + ρτ Ψ22 (99)

Thus, the minimum size of Br that ensures this condition has radius given by (97).

Remark 4.5. The proofs of Theorem 4.1 and Corollary 4.1 assume that the sets

Dx and Dq are sufficiently large. If we define Br∗ as the largest ball contained in Dq ,

61
and assume that the initial conditions are such that q(0) ⊂ Br∗ , then from Figure 1
we have the added condition that r < r ∗ , which implies a lower bound on ρ. It can

be shown that in this case the lower bound must be such that λmin (P̃ ) = ρ. With r

defined by (97) and λmin (P̃ ) = ρ, the condition r < r ∗ implies:

λ(P )Ψ21
ρ > ∗2 (100)
r − τ Ψ22
Since κ2 = 1/ρη, η > 1, it follows from (100) that r ∗ should ensure that:

r ∗2 − τ Ψ22
κ2 < (101)
λ(P )Ψ21
Therefore, the meaning of Dq sufficiently large in Corollary 4.1 is that
q
∗
r > κ2 λmax (P )Ψ21 + τ Ψ22 (102)

and q(0) ⊂ Dr∗ . The meaning of Dx sufficiently large is difficult to characterize

precisely since x(t) depends on both r(t) and x(0). Nevertheless it can be seen that

increasing κ2 implies increasing the require size of the set Dx .

Corollary 4.2. Under the conditions of Theorem 4.1, the error trajectory ap-

proaches the ultimate bound exponentially in time according to

|q(t)| ≤ k|q(0)|e−cr t , t<T (103)

with a convergence rate given by:

τ /(1 + τ )Ψ21 ρ(η − 1)

cr = , τ= (104)
2λmax (P̃ ) λmin (Qm )
Proof. Choosing ξ so that c2 = c1 τ , and using the expressions for c1 and c2 in

(88) and (89) results in the expression for τ in (104). Then, from (87), we can write:

V̇(e(t), W̃t ) ≤ −c1 |q(t)|2 + d (105)

p
Define ĉ ≡ d/µ2 , where µ ≡ Ψ21 + τ Ψ22 . Then, when |q(t)| > µ, we have that

V̇(e(t), W̃t ) ≤ −(c1 − ĉ)|q(t)|2

≤ −k3 |q(t)|2 (106)

62
where k3 ≡ c1 − ĉ. Now, c1 = d/Ψ21 and c2 = d/Ψ22 , and since c2 /c1 = τ , it follows
that Ψ21 = τ Ψ22 , and therefore ĉ = d/(1 + τ )Ψ21 . Therefore k3 = τ /(1 + τ )Ψ21 . Finally,

since k1 |q(t)|2 ≤ V(e(t), W̃t ) ≤ k2 |q(t)|2 , where k1 = λmin(P̃ ) > 0, k2 = λmax (P̃ ) > 0,

and V̇(e(t), W̃t ) ≤ −k3 |q(t)|2, then (103) follows directly from Corollary 5.3 of Ref. 45
p q
where k ≡ k2 /k1 = λmax (P̃ )/λmin(P̃ ), and cr = (k3 /2k2 ) = τ /2(1 + τ )λmax (P̃ )Ψ21 .

Remark 4.6. For the most common case in which λmax (P̃ ) > ρ, there exists:
q
∗
η = (−ϕ1 + ϕ21 − 4ϕ2 )/2 (107)
2
ϕ1 , − (108)
1 + κ2 λmin (Qm )
κ1 − κ2 λmin(Qm )
ϕ2 , (109)
κ1 (1 + κ2 λmin (Qm ))

where η ∗ is the the value of η such that the convergence rate cr in (104) attains a

maximum on the open interval (1, 1/κ1).

Figures 32 and 33 show plots of the normalized ultimate bound (r/δ ∗ ) in (97)

versus convergence rate cr in (104) for λmax (P̃ ) = 2, and λmin (P̃ ) = λmin(Qm ) = 1,

Figure 32: r/δ ∗ versus cr for κ1 = {0.01, 0.02, 0.03} and κ2 ∈ [1, 10].

63
 Figure 33: r/δ ∗ versus cr for κ1 = {0.1, 0.2, 0.3} and κ2 ∈ [1, 10].

and η = η ∗ . Figure 32 shows that increasing κ2 increases the ultimate bound, but

it also increases the convergence rate. The fact that the estimate for the ultimate

bound increases with the adaptation gain has been previously noted in Refs. 12 and

83. Figure 32 also shows that increasing κ1 from 0.01 to 0.03 reduces the ultimate
bound for any given value of κ2 ∈ [1, 10]. However, if we increase κ1 from 0.1 to 0.3,

then Figure 33 shows that the trend in Figure 32 reverses for values κ2 ∈ [1, 10]. This

calculation can be repeated for any given set of parameters in (97) and (103) as a

guide to choosing the design parameters of DF-MRAC.

For the case of constant ideal weights in Assumption 4.1, Theorem 4.1 specializes

to the following theorem. In this case, we assume that the uncertainty is structured.
That is, the vector of known functions in (74) represent the vector of known basis

functions.

Theorem 4.2. Consider the uncertain system given by (1) subject to Assumption

4.1, where W ∈ IRs×m is an unknown constant weight matrix. Consider, in addition,

the feedback control law given by (5), with the nominal feedback control component

64
given by (2) subject to Assumption 2.2, and with the adaptive feedback control com-
ponent given by (6) that has the derivative-free weight update law in the form (76)

and (78), where Ω1 = I. Then, e(t) and W̃ (t) approach a subspace in these error

variables in which e(t) = 0 and B W̃ T (t)β(x(t)) = 0.

Proof. The result follows directly from the proof of Theorem 4.1 by choosing

Ω1 = I. In this case, δ ∗ = 0 due to Ω2 (t) = 0 in (79), which follows from the fact that

the ideal weights are constant, and Ω1 = I. Then, the inquality in (85) is not needed
1
since the left hand side vanishes. In this case, κ1 = η
= 1, ξ = 0, c2 = 0 in (89),

c3 = 0 in (90), and d = 0 in (91). Therefore, it follows from (87) that the entire error

space is invariant. Let E denote the set of points in this space where V̇(e(t), W̃t ) = 0.

From (87) all points in lie in a subspace, where e(t) = 0 and it follows from LaSalle’s

Theorem [45] that all solutions in the error space approach the largest invariant set

M in E. Now e(t) = 0 implies that Ω̂2 (t) = 0 and ė(t) = 0. Then (81), with all of the
above taken together, implies that M is comprised of all points in the error space in

which e(t) = 0 and B W̃ T (t)β(x(t)) = 0. 

Remark 4.7. The system is said to be sufficiently excited if r(t) is such that the

conditions e(t) = 0, B W̃ T (t)β(x(t)) = 0 admit only the solution W̃ (t) = 0 in the

limit t → ∞. It is straightforward to show that this amounts to the standard MRAC

condition for persistency of excitation [71].

Remark 4.8. As noted in the introduction, DF-MRAC does not employ an

integrator in its weight update law. This is advantageous from the perspective of

augmenting a nominal controller that employs integral action to ensure that the

regulated output variables track r(t) for constant disturbances, regardless of how

these disturbances may enter the system. An example that illustrates this advantage

is provided in Section 4.6.2.

Remark 4.9. The expressions for Ψ1 and Ψ2 represent ultimate bounds for |e(t)|

and ||W̃ (t)||, respectively. These expressions depend on η > 1. In Remark 4.6 it

65
is shown that there exists an optimal value η ∗ such that the convergence rate cr in
(104) attains a maximum on the open interval {1, 1/κ1 }, and curves are provided

that show the trade-off between the UUB and cr for a range of values for κ1 and κ2 .

4.4 Modifications to Derivative-Free Adaptive Control

Although the derivative-free weight update law does not require a modification term

to prove the error dynamics are UUB, one may wish to employ a modification term

in order to improve performance or robustness of the system. The following theorem

extends Theorem 4.1 to a general form of modified DF-MRAC.

Theorem 4.3. Consider the uncertain system given by (1) subject to Assumption

4.1. Consider, in addition, the feedback control law given by (5), with the nominal

feedback control component given by (2) subject to Assumption 2.2, and with the

adaptive feedback control component given by (6) that has a derivative-free weight

update law in the form given by (76), where τ > 0 is a time delay design value, and

Ω1 ∈ IRs×s and Ω̂2 (t) ≡ Ω̂2 (x(t), e(t)), Ω̂2 : IRn × IRn → IRs×m , satisfy:

0 < ΩT
1 Ω1 < κ1 I, 0 < κ1 < 1 (110)

Ω̂2 (t) = κ2 [β(x(t))eT (t)P B − κm S(t)Ŵ (t)], κm > 0 (111)

where κm is the modification gain, P ∈ IRn×n satisfies

0 = AT 2 T
m P + P Am − κm P BB P + Qm (112)

for any symmetric Qm > 0, and S(t) ∈ IRs×s satisfies ||S(t)|| < s∗ . Then, e(t) and

W̃ (t) are UUB.

Proof. The result follows directly from the proof of Theorem 4.1, with (88) - (91)

66
changed to:

c1 = λmin(Qm ) > 0 (113)

2
c2 = ρξ + s∗ κ∗ > 0 (114)

c3 = ρλmin (I − [1 + η + γ]ΩT
1 Ω1 ) > 0 (115)
2 2 2
d = ρ(1 + η + η 2 /γ)δ ∗ + ω ∗ s∗ κ∗ > 0 (116)

where κ∗ = 1 + 2κm + (1 + ξ)−1 κm 2 > 0. 

Remark 4.10. If S(t) is not assumed to be bounded by s∗ , e.g. S(t) = ||e(t)|| in

the case of e- modification [37], then one can show that e(t) and W̃ (t) are UUB by

applying a projection operator [66] to the weight estimates given by (76).

Remark 4.11. In the case of σ- modification [60], we have S(t) = I in (111).

ALR modification [10] can be introduced by letting S(t) = βx (x(t))βxT (x(t)), where

βx (x(t)) , dβ(x(t))/dx(t) ∈ IRs×n and in this case s∗ = 1. Table 1 summarizes DF-

MRAC laws for these two and other modification terms as well. Note that since Ŵ (t)
appears in both left and right hand sides of some DF-MRAC laws with modification

terms, then one needs to solve these expressions for Ŵ (t) for implementation.

67
 Table 1: DF-MRAC laws for various modification terms

Modification DF-MRAC Law for −1 < ϕ1 < 1, ϕ2 > 0, ϕ̂2 > 0

 
Original (76) Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B .
 
σ [60] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B − ϕ̂2 Ŵ (t) .
 
e [37] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B − ϕ̂2 ke(t)kŴ (t) .
 
ALR [10] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B − ϕ̂2 βx (x(t))βxT (x(t))Ŵ (t) .


Q [75, 77] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B − ϕ̂2 Ŵ T q(t) − c(t) q(t) ;
Rt
where q(t) , td β(x(s))ds, td > 0, c(t) , en (t) − en (td )
Rt Rt
− td B T Am × Am e(s)ds + td uad (s)ds, and en (t) is the n−th
component of the vector e(t), and td > 0.

Optimal [63] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B + ϕ̂2 β(x(t))β(x(t))T

×Ŵ (t)B T P A−1m B .
 
CMRAC [50] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B − ϕ̂2 βf (x(t))eT
Y (t) ;
λf
where βf (x(t)) = L−1 { s+λ f
}β(x(t)), where L−1 denotes
the Laplace operator, and eY (t) is the predictor estimation error
defined in Ref. 50.
 Rt 
K [47] Ŵ (t) = ϕ1 Ŵ (t − τ ) + ϕ2 β(x(t))eT P B − ϕ̂2 t−td Ŵ (s)ds ;
where td > 0.

68
4.5 Extensions to the Input Uncertainty Case

Consider the system with matched uncertainty and input uncertainty of the form

ẋ(t) = Ax(t) + BΛu(t) + B∆(t, x(t)) (117)

where Λ ∈ IRm×m is an uncertain diagonal matrix with diagonal elements λi ≥ 0.

Assumption 4.2. R{B} = R{BΛ}.

Remark 4.12. Assumption 4.2 is automatically satisfied when λi > 0, therefore

it is only of importance in situations wherein one or more of the λi = 0 corresponds

to actuator failure. In this case Assumption 4.2 implies that there exists actuator

redundancy. An actuator is considered redundant if its failure does not change the

range space of B. This is approximately true in aircraft applications where the means

of control is primarily through moment generation. Therefore, Assumption 4.2 is

approximately satisfied when there is sufficient actuator redundancy. For situations

in which there is no actuator redundancy, and one or more λi = 0, it may still be

possible to control the aircraft through a hierarchal adaptive approach [35], which

does not require Assumption 4.2. In Section 4.7 we illustrate the application of

Assumption 4.2 for the case of redundant control. The manner in which redundant

controls can be implemented is also illustrated in Section 4.7.

Under Assumption 4.2, the dynamics (117) can equivalently be written in the

form
 
ẋ(t) = Ax(t) + BΛ u(t) + ∆e (t, x(t)) (118)

where ∆e (x(t)) also satisfies Assumption 4.1. The state tracking error dynamics in

(75) becomes
   
ė(t) = Am e(t) + B Λ − Im K1 x(t) + B Λ − Im K2 r(t)
 
+BΛ −uad (t) + W T (t)β(x(t)) (119)

69
Assumption 4.2 allows us to write (119) in the equivalent form
 
ė(t) = Am e(t) + BΛ −uad (t) + W T (t)β(x(t)) + K1e
T T
x(t) + K2e r(t) (120)

The adaptive law is now chosen in the form

uad (t) = Ŵ T (t)β(x(t)) + K̂1e

T T
(t)x(t) + K̂2e (t)r(t) (121)

Using (121) in (120) yields

 
ė(t) = Am e(t) + BΛ W̃ T (t)β(x(t)) + K̃1e
T T
(t)x(t) + K̃2e (t)r(t) (122)

where

K̃1e (t) ≡ K1e − K̂1e (t) (123)

K̃2e (t) ≡ K2e − K̂2e (t) (124)

The following theorem states an extension to Theorem 4.1 for the case of input un-

certainty.

Theorem 4.4. Consider the uncertain system given by (117) subject to Assump-

tions 4.1 and 4.2. Consider, in addition, the feedback control law given by (5), with
the nominal feedback control component given by (2) subject to Assumption 2.2, and

with the adaptive feedback control component given by (121), that has derivative-free

weight update laws in the form given by (76), (77), (78), and

K̂1e (t) = Ξ11 K̂1e (t − τ ) + Ξ̂12 (t) (125)

K̂2e (t) = Ξ21 K̂2e (t − τ ) + Ξ̂22 (t) (126)

where τ > 0, and Ξ11 ∈ IRn×n , Ξ21 ∈ IRr×r , Ξ̂12 : IRn × IRn → IRn×m , and Ξ̂22 :

IRn × IRn → IRr×m satisfy:

0 ≤ ΞT
i1 Ξi1 < I, i = 1, 2 (127)

Ξ̂12 (t) = κ2 x(t)eT (t)P B, κ2 > 0 (128)

Ξ̂22 (t) = κ2 r(t)eT (t)P B (129)

70
with P ∈ IRn×n satisfying (8) for any symmetric Qm > 0. Then, e(t), W̃ (t), K̃1e (t),
and K̃2e (t) are UUB.
 T T
  T
Proof. Let We (t) ≡ W T (t), K1e T
, K2e , Ŵe (t) ≡ Ŵ T (t), K̂1e
T T
(t), K̂2e (t) , W̃e (t)
 T
≡ We (t) − Ŵe (t), and βe (β(x(t)), x(t), r(t)) ≡ β T (x(t)), xT (t), r T (t) . Then, the

state tracking error dynamics in (122) becomes:

ė(t) = Am e(t) + BΛW̃eT (t)βe (β(x(t)), x(t), r(t)) (130)

Defining

Ω2e (t) ≡ We (t) − Ω1e We (t − τ ) (131)

where Ω1e ≡ diag[Ω1 , Ξ11 , Ξ21 ] and ||Ω2e (t)|| ≤ δe∗ , W̃e (t) can be written as

W̃e (t) = Ω1e W̃e (t − τ ) + Ω2e (t) − Ω̂21 (t) (132)

where Ω21 (t) ≡ [Ω̂T T T T

2 (t), Ξ12 (t), Ξ22 (t)] . The state tracking error dynamics in (130)

can be rewritten using (132) as:

 T
ė(t) = Am e(t) + BΛ Ω1e W̃e (t − τ ) + Ω2e (t) − Ω̂21 (t) (t)βe (·) (133)

To show that the closed-loop system given by (132) and (133) is UUB, consider

the Lyapunov-Krasovskii functional

Z t
T


V(e(t), W̃et ) = e (t)P e(t) + ρ tr W̃eT (s)W̃e (s) Λ + εIm ds (134)
t−τ

where ρ > 0, W̃et represents W̃ (t) over the time interval t − τ to t, and ε > 0 is

sufficiently small. By following identical steps presented in the proof of Theorem 4.1,

the directional derivative of (134) along the closed-loop system trajectories of (132)

and (133) is given by

V̇(e(t), W̃et ) ≤ −c̄1 ke(t)k2 − c̄2 kW̃e (t)k2 − c̄3 kW̃e (t − τ )k2 + de (135)

71
where the constants c̄1 , c̄2 , c̄3 , and d¯ are:

c̄1 = λmin(Qm ) > 0 (136)

c̄2 = ρϑξ > 0 (137)

c̄3 = ρϑλmin (I − ΩT
1e Ω1e ) > 0 (138)
2
de = ρϑ(η + η 2 /γ)δe∗ > 0 (139)

where ϑ = λmin(Λ + εIm ) and ϑ = λmax (Λ + εIm ). Either |e(t)| > Ψ̄1 or kW̃e (t)k > Ψ̄2
p p
or kW̃e (t − τ )k > Ψ̄3 renders V̇(e(t), W̃et ) < 0, where Ψ̄1 = de /c̄1 , Ψ̄2 = de /c̄2 ,
p
and Ψ̄3 = de /c̄3 . Hence, it follows that e(t) and W̃e (t) are UUB. 

Remark 4.13. For the simplicity of presentation, Theorem 4.4 assumes that the

derivative-free weight update laws given by (76), (125), and (126) use the same τ > 0

and κ2 > 0 design parameters. To add greater flexibility in the design, (125), (126),

(128), and (129) can be replaced by:

K̂1e (t) = Ξ11 K̂1e (t − τ1e ) + Ξ̂12 (t), τ1e > 0 (140)

K̂2e (t) = Ξ21 K̂2e (t − τ2e ) + Ξ̂22 (t), τ2e > 0 (141)

Ξ̂12 (t) = κ12 x(t)eT (t)P B, κ12 > 0 (142)

Ξ̂22 (t) = κ22 r(t)eT (t)P B, κ22 > 0 (143)

respectively. In this case, the proof of Theorem 4.4 follows a similar line by considering
Z t
T


V(e(t), W̃et ) = e (t)P e(t) + ρ tr W̃ T (s)W̃ (s) Λ + εIm ds
t−τ
Z t
 T


+ρ1e tr K̃1e (s)K̃1e (s) Λ + εIm ds
t−τ
Z 1e
 t T


+ρ2e tr K̃2e (s)K̃2e (s) Λ + εIm ds (144)
t−τ2e

where ρ1e > 0 and ρ2e > 0.

Remark 4.14. Under the conditions of Theorem 4.4, an estimate for the ultimate

bound and convergence rate to this ultimate bound can be expressed in a form similar

72
to Corollaries 4.1 and 4.2, respectively. For the case of constant ideal weights in
Assumption 4.1, Theorem 4.4 with Ω1 = I, Ξ11 = I, and Ξ21 = I guarantees that

e(t), W̃ (t), and D̃(t) approach a subspace in these error variables in which e(t) = 0

and BΛW̃eT (t)βe (·) = 0, similar to the result given by Theorem 4.2. Modifications

to the derivative-free weight update laws in (76), (125), and (126) can be employed,

analogous to given in Table 1.

4.6 Examples on a First Order System

In this section we compare the standard MRAC law given by (7) with the proposed
DF-MRAC law given by (76) on a simple model for the rolling dynamics of an aircraft

[53].

4.6.1 Uncertainty with Time-Varying Ideal Weights

Consider the scalar dynamics

ẋ(t) = Lp x(t) + Lδ [u(t) + ∆(t, x(t))] (145)

where x(t) represents roll rate, u(t) represents aileron deflection, Lp = −1, Lδ = 1,

and ∆(x(t)) = w(t)x(t) with w(t) being a square wave having an amplitude of 1.0,

and a period of 5.0 seconds. While aircraft dynamics do not behave this manner,

their stability derivatives can undergo a sudden change in the event of damage to the

airframe. So this example should be regarded in this context. In both the MRAC
and DF-MRAC architectures, we choose as a reference model Am = −2 and Bm = 2.

Note that for this example ε(x) = 0, so we may use the MRAC law given by (7)

without a modification term, and choose β(x) = x. Furthermore, we consider two

adaptation gains, γ = 102 and γ = 104 . For the DF-MRAC law given by (76), we set

τ = 0.01 seconds, Ω1 = 0.5 which satisfies (77) , and κ2 = 200 in (78). Qm = 2 was

used in (8) to solve for P in both architectures.

73
 Remark 4.15. It is important to mention that we implemented standard MRAC
in analog form. Therefore, all discrete time implementations used to realize the

integration (see, for example, Remark 4.3) will produce nearly the same results, since

they all represent an approximation to an integration performed in continuous time,

and then sampled at the control update rate.

Figure 34 shows the performance of the nominal control design without adap-

tation. In Figure 35, the standard MRAC architecture is used with γ = 102 and
γ = 104 , respectively. Tracking performance is not improved by increasing the adap-

tation gain beyond 102 , and increasing gain causes high frequency oscillations in the

control response that would be unacceptable in a real system. Figure 36 shows the

case in which the DF-MRAC adaptive law in (76) was used with κ2 = 200. It can be

seen that tracking performance is excellent, and the control time history is reasonable,

and in this case the estimated weight is close to the ideal value.
As a variation of the previous example, the ideal weight history was changed

to a sinusoidal function in which the frequency was varied from 0.5 rad/s up to 50

rad/s. The results are shown in Figures 37–39. In Figure 37, note that the adaptive

controller does not significantly improve the response at the low setting for adaptation

gain, and gives an even worse response when using the high setting for adaptation

gain. Figures 38 and 39 show that the DF-MRAC case gives an excellent response,
and the estimated weight is very close to the true weight, even at the high frequency

end (see Figure 39). Inspired by this result we decided to try a case in which w(t)

is a band-limited white noise signal. The associated results are shown in Figures 40

and 41. In particular, Figure 41 shows that the estimated weight is remarkably close

to to w(t).

74
 2
r(t)
1 x(t)

r(t), x(t)
0

−1

−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)
4
u(t)
u(t) 2

−2

−4
0 2 4 6 8 10 12 14 16 18 20
t (sec)
1
ideal weight
0.5 estimated weight
weight

−0.5

−1
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Figure 34: Responses with nominal controller for the square wave ideal weight.

1
r(t), x(t)

0
r(t)
x(t) (MRAC, γ=102)
−1
x(t) (MRAC, γ=104)
−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)
6
4
2
u(t)

0
−2 u(t) (MRAC, γ=102)
−4 u(t) (MRAC, γ=104)
−6
0 2 4 6 8 10 12 14 16 18 20
t (sec)
3
2
1
weight

0 ideal weight
−1 estim. weight (MRAC, γ=102)
−2 estim. weight (MRAC, γ=104)
−3
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Figure 35: Responses with standard MRAC using γ = 102 and γ = 104 .

75
 1 r(t)
x(t)
0.5

r(t), x(t)
0
−0.5
−1

0 2 4 6 8 10 12 14 16 18 20
t (sec)
5
u(t)
u(t)

−5
0 2 4 6 8 10 12 14 16 18 20
t (sec)
2
ideal weight
1 estimated weight
weight

−1

−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Figure 36: Responses with DF-MRAC for the square wave ideal weight.

1
r(t), x(t)

0 r(t)
2
x(t) (MRAC, γ=10 )
−1
x(t) (MRAC, γ=104)
−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)
20

10
u(t)

0
2
u(t) (MRAC, γ=10 )
−10
u(t) (MRAC, γ=104)
−20
0 2 4 6 8 10 12 14 16 18 20
t (sec)
10

5
weight

0
ideal weight
−5
estim. weight (MRAC, γ=102)
−10 4
estim. weight (MRAC, γ=10 )
−15
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Figure 37: Responses with standard MRAC using γ = 102 and γ = 104 .

76
 2
r(t)
1 x(t)

r(t), x(t)
0

−1

−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)
6
4 u(t)
2
u(t)

0
−2
−4
−6
0 2 4 6 8 10 12 14 16 18 20
t (sec)
4
ideal weight
2 estimated weight
weight

−2

−4
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Figure 38: Responses with DF-MRAC for a sinusoidal ideal weight.

ideal weight
3 estimated weight

1
weight

−1

−2

−3

18.5 19 19.5 20
t (sec)

Figure 39: Expanded view of the estimate of the ideal weight.

77
 2
r(t)
1 x(t)

r(t), x(t)
0

−1

−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)
20
u(t)
10
u(t)

−10

−20
0 2 4 6 8 10 12 14 16 18 20
t (sec)
20
ideal weight
10 estimated weight
weight

−10

−20
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Figure 40: Responses with DF-MRAC for a band limited white noise ideal weight.

12
ideal weight
estimated weight
10

2
weight

−2

−4

−6

−8

−10
18.5 19 19.5 20
t (sec)

Figure 41: Expanded view of the estimate of the ideal weight.

78
4.6.2 Uncertainty with Constant Ideal Weights

Consider the scalar dynamics given by (145) with ∆(t, x) = 1 + 5x. To illustrate the

point that conventional MRAC is problematic when augmenting a nominal controller

that has integral action, we introduce an integrator state

ẋi (t) = −x(t) + r(t) (146)

and consider the augmented dynamics

     
 Lp 0   Lδ   0 
   
ẋa (t) =   xa (t) +   u(t) + 1 + 5x(t) +   r(t) (147)
−1 0 0 1

where xT
a (t) = [x(t) xi (t)], and choose a PI form for the nominal controller:

un (t) = −[kx ki ]xa (t) (148)

Then, the reference model in (4) becomes:

   
 Lp − Lδ kx −Lδ ki   0 
ẋm (t) =   xm (t) +   r(t) (149)
−1 0 1

The adaptive control is chosen in the form of (6), with β T (x(t)) = [1 xT

a (t)], and uses

the same weight update law as in the previous example. We again let Lp = −1 and

Lδ = 1, and the PI gains are set to kx = 1.5 and ki = 3. For the MRAC law given

by (7), we consider two adaptation gains, γ = 20 and γ = 200. For the DF-MRAC

law, the parameter settings are the same as in the previous example. Figures 42
and 43 present results that support the statement made in Remark 4.8. Figure 42

illustrates the conflict that can arise when a nominal control law with integral action

is augmented with an adaptive controller that employs an integrator in its weight

update law. Note the slowly varying tracking error during the first transient phase.

In more complex problems this slow variation can be persistent. Figure 43 shows that

the DF-MRAC law works well, even during the first transient phase.

79
 2
r(t)
1 x(t) (MRAC, γ=20)

r(t), x(t)
x(t) (MRAC, γ=200)
0

−1

−2
0 5 10 15
t (sec)
5
u(t) (MRAC, γ=20)
0 u(t) (MRAC, γ=200)
u(t)

−5

−10
0 5 10 15
t (sec)
15
∆(t)
10 estim. ∆(t) (MRAC, γ=20)
estim. ∆(t) (MRAC, γ=200)
∆(t)

−5
0 5 10 15
t (sec)

Figure 42: Responses with standard MRAC for an uncertainty with constant ideal
weights.

2
r(t)
1 x(t)
r(t), x(t)

−1

−2
0 5 10 15
t (sec)
5
u(t)
0
u(t)

−5

−10
0 5 10 15
t (sec)
10
∆(t)
5 estim. ∆(t)
∆(t)

−5
0 5 10 15
t (sec)

Figure 43: Responses with DF-MRAC for an uncertainty with constant ideal weights.

80
4.6.3 Input Uncertainty Case

Consider the scalar dynamics

ẋ(t) = Lp x(t) + Lδ [Λu(t) + ∆(t, x(t))] (150)

where Λ = 0.1 represents an input uncertainty. Let w(t) be a sinusoidal time-varying

ideal weight with an amplitude of 1 and a frequency of 2.5 seconds. Furthermore,

consider the same adaptive control design in Section VI.A for (76) and choose Ξ11 =

Ξ21 = 0.5 for (125) and (126). Figure 44 shows the result when (76) is employed

and Figure 45 shows the result when (76), (125), and (126) are employed. There is a

dramatic improvement when (76), (125), and (126) are employed.

1
r(t)
x(t)
0.5 x (t)
m
r(t), x(t), x (t)
m

−0.5

−1
0 5 10 15 20 25 30
t (sec)

20

10
u(t)

−10

−20

0 5 10 15 20 25 30
t (sec)

Figure 44: Responses with DF-MRAC for input uncertainty case when (76) is em-
ployed.

81
 1
r(t)
x(t)
0.5 x (t)
m
r(t), x(t), xm(t)

−0.5

−1
0 5 10 15 20 25 30
t (sec)

20

10
u(t)

−10

−20

0 5 10 15 20 25 30
t (sec)

Figure 45: Responses with DF-MRAC for input uncertainty case when (76), (125),
and (126) are employed.

82
4.7 Examples on a Generic Transport Model

This section presents a DF-MRAC design for the NASA GTM, and evaluates this
design for several damage scenarios. GTM is a high-fidelity scaled transport aircraft

model developed by NASA Langley Research Center [1]. A linearized model for GTM

at an angle of attack of 2 degrees and 104 ft altitude is obtained in the form of (1).

The primary sources of uncertainty are any one of a set of possible damage conditions

that are included as a part of the modeling in GTM.

A nominal controller is first designed for the linearized model using a robust
servomechanism LQR approach that incorporates integral control [79], with the ob-

jective of tracking roll rate, pitch rate, and yaw rate commands. Including the in-

tegral states, the linearized GTM model is 9th -order with the state vector x(t) =
 T
u(t) v(t) w(t) p(t) q(t) r(t) φ(t) θ(t) qint (t) pint (t) wint (t) , where u(t), v(t), w(t) are

velocity components; p(t), q(t), r(t) are body angular rates about the roll, pitch and

yaw body axes; φ(t) and θ(t) are roll and pitch attitude; and qint (t), pint (t), wint (t) are
the integrator states. In this simulation study, tracking of roll and pitch rate com-

mands are considered, and yaw rate command is set to zero. Roll and pitch attitude

are not used in the design. Figure 46 shows the performance of the nominal controller

under normal operating conditions.

Since this design has redundant actuation and damage conditions that may include

loss of one or more actuator channels, it is necessary to generalize the form in (117)
to
 
ẋ(t) = Ax(t) + BΛ Gun (t) − uad (t) +B∆(t, x(t)) (151)

uad (t) = Ŵ T (t)β(x(t)) + GK̂1e

T T
(t)x(t) + GK̂2e (t)r(t) (152)

where G ∈ IRM ×3 is a control allocation matrix in which M > 3 denotes the number

of independent control channels, un ∈ IR3 is the effective nominal control for the roll,

83
pitch, and yaw axes, respectively, and ua ∈ IRM is the adaptive control. Note that
while the nominal control law must operate through the control allocation matrix,

a portion of the adaptive controller has direct access to each independent channel

of actuation. The quantity Λ ∈ IRM ×M is nominally an identity matrix, and loss of

actuation is represented by setting one or more of its diagonal elements to zero, and

∆ : [0, ∞) × IRn → IRn is the uncertainty, which primarily enters the p, q, and r

state equations. It is used to represent uncertainty in the stability derivatives. In

general the portion of the uncertainty that remains matched under actuator failure

corresponds to the projection of ∆ onto the column space of BΛ. For aircraft flight

control applications the assumption that ∆ remains matched under actuator failure

amounts to assuming that BΛ and ∆ primarily influence the moments acting on the

aircraft, and that control of moments in all three axes is maintained under actuator

failure. For this study only the spoilers are independently controlled, so there is a
total of M = 6 independent control channels: elevator, aileron, rudder, left and right

spoilers, and stabilizer. Since there are 6 independent control surfaces and only 3

moment axes, it is conceivable that up to 3 control surfaces could fail with minimal

change in the column space of B. In general, this is approximately true in aircraft

applications where the means of control is primarily through moment generation.

Therefore, Assumption 4.2 is not restrictive for these cases when we have actuator
redundancy. On the other hand, for the cases when there is no actuator redundancy,

this is a restrictive assumption. Servo dynamics and position and rate limits are

included in the GTM model. The stabilizer servo has a relatively low bandwidth and

low value for its rate limit, and is useful primarily for maintaining trim in the pitch

axis. The nominal control design is performed using BG in place of B in (1) when

doing the design.

For the adaptive design, neural network sigmoidal type functions are used in the

form β(x) = [1, β1 (x), β2 (x), . . . , β6 (x)]T , where βi (x) = (1 − e−xi )/(1 + e−xi ), i =

84
 
1, . . . , 6, and P in (8) is computed using Qm = diag 10−4 10−2 10−3 60 30 15 10 10 10].
Taking into consideration Theorem 4.4, weight update laws with ALR modification

are:
h 
Ŵ (t) = Ω1 Ŵ (t − τ ) + β(x(t))eT P B − κm βx (x(t))βxT (x(t))Ŵ (t)
i
+βx (x(t))K̂1e (t)GT Γκ2 , Γκ2 > 0 (153)

 
K̂1e (t) = Ξ11 K̂1e (t − τ1e ) + κ21 x(t)eT (t)P B (154)
 
K̂2e (t) = Ξ21 K̂2e (t − τ2e ) + κ22 r(t)eT (t)P B (155)

 
We chose Ω1 = 0.1, Γκ2 = diag 10 1 10 10 50 10 , and τ = 0.5 for (153), and

Ξ11 = 0.1I3 , Ξ21 = 0.1I2 , κ12 = κ22 = 0.05, and τ1e = τ2e = 0.5, for (154) and (155).
The ALR modification gain in (153) was chosen as κm = 5.0.

Remark 4.16. The form of the ALR modification term in (153) is different from

the form given in Table 1. Following the procedure in Ref. 10, the form in Table 1

is obtained by taking the gradient of ||∂uad (Ŵ , x)/∂x||2 /2 with respect to Ŵ , with

uad (Ŵ , x) defined by (6), whereas the form in (153) is obtained with uad (Ŵ , x) defined

by (152).
Remark 4.17. The form of the adaptation gain, Γκ2 > 0, in (153) is more

general that the scalar form used in (78). This was done to add greater flexibility in

the design. The proof of Theorem 4.1 can be generalized to this case using:
Z t
T
 
V(e(t), W̃t ) = e (t)P e(t) + tr W̃ T (s)W̃ (s)Γρ ds , Γρ > 0 (156)
t−τ

4.7.1 Missing Left Wing Tip

In the missing left wing tip damage scenario there is 25% loss of outboard left wing

tip and the left aileron is missing, therefore available roll control effectiveness is

reduced [1]. Figure 47 shows the degree of degraded performance when the nominal

controller is evaluated for this damage case. Figure 48 shows the improvement in

85
response obtained when DF-MRAC is employed. In Figure 49, 0.01 seconds of time
delay is added to the rudder channel. In this case, the performance using DF-MRAC

alone is not satisfactory. Figure 50 shows that the addition of ALR modification

significantly improves the robustness of the adaptive controller to time delay.

8 8
p q
6 cmd 6 cmd

4
p 4
q
p [deg/sec]

q [deg/sec]
2 2

0 0

−2 −2

−4 −4

−6 −6

−8 −8
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

2.5 3.5
r
cmd 3
α
2
r β
2.5
r [deg/sec]

1.5

α, β [deg]
2

1 1.5

1
0.5
0.5
0
0

−0.5 −0.5
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

5.5 0.8 0.12 1 1

5
left elev. 0.6
left ail. 0.1
rud. left spl. stab.
right elev. right ail. right spl.

stabilizer [deg]
4.5 0.4 0.08 0.5 0.5
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]

4 0.2 0.06

3.5 0 0.04 0 0

3 −0.2 0.02

2.5 −0.4 0 −0.5 −0.5

2 −0.6 −0.02

1.5 −0.8 −0.04 −1 −1

0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 46: GTM nominal control response for nominal operating conditions.

86
 8
p q
20 cmd 6 cmd
p q
4
10

p [deg/sec]

q [deg/sec]
2
0
0
−10 −2

−20 −4

−6
−30
−8
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

4 6
r
cmd
α
5
2
r β
4
r [deg/sec]

α, β [deg]
3
−2
2
−4
1

−6 0

−8 −1
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

6 0 0.8 1 1
left elev. left ail. rud. left spl. stab.
5
right elev. right ail. 0.6 right spl.

stabilizer [deg]
−5 0.5 0.5
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]
4

0.4
3
−10 0 0
2
0.2

1
−15 −0.5 −0.5
0
0

−1 −20 −0.2 −1 −1
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 47: GTM nominal control response for the missing left wing tip case.

8
pcmd qcmd
20 6
p q
4
10
p [deg/sec]

q [deg/sec]

2
0
0
−10 −2

−20 −4

−6
−30
−8
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

3 7
r
cmd 6
α
2
r β
5
r [deg/sec]

1
α, β [deg]

0 3

2
−1
1
−2
0

−3 −1
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

5 1 1.5 25 0.3
left elev. 0
left ail. rud. left spl. stab.
4 right elev. right ail. right spl.
1 20 0.2
stabilizer [deg]

−1
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]

3
−2 0.5 15 0.1
2 −3

−4 0 10 0
1
−5
−0.5 5 −0.1
0
−6

−1 −7 −1 0 −0.2
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 48: GTM DF-MRAC response for the missing left wing tip case.

87
 40 15
p q
cmd cmd
20 p 10 q

p [deg/sec]

q [deg/sec]
0 5

−20 0

−40
−5

−60
−10
0 2 4 6 8 10 12 0 2 4 6 8 10 12
time [sec] time [sec]

60
r
cmd
6
α
40
r 4 β
r [deg/sec]

α, β [deg]
20
2

0 0

−20 −2

−40 −4

0 2 4 6 8 10 12 0 2 4 6 8 10 12
time [sec] time [sec]

40
left elev. left ail. 10 rud. left spl. 0.2 stab.
4 35
right elev. 10 right ail. right spl.
0.15

stabilizer [deg]
elevator [deg]

30
aileron [deg]

spoiler [deg]
rudder [deg]
3 5
5 25 0.1

2 0 20
0.05
0 15
−5 0
1 10
−5 −0.05
−10 5
0
0 −0.1
0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 49: GTM DF-MRAC response for the missing left wing tip case with 0.01
seconds of time delay in the rudder channel.

8
p q
20 cmd 6 cmd
p q
4
10
p [deg/sec]

q [deg/sec]

2
0
0
−10 −2

−20 −4

−6
−30
−8
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

3 8
rcmd α
2
r 6 β
r [deg/sec]

1
α, β [deg]

4
0
2
−1

0
−2

−3 −2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

5 1 2.5 25 0.3
left elev. left ail. rud. left spl. stab.
4 0 2
right elev. right ail. 20 right spl. 0.2
stabilizer [deg]
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]

1.5
3 −1
15 0.1
1
2 −2
0.5
10 0
1 −3
0
5 −0.1
0 −4 −0.5

−1 −5 −1 0 −0.2
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 50: GTM DF-MRAC response with ALR modification term for the missing
left wing tip case with 0.01 seconds of time delay in the rudder channel.

88
4.7.2 Missing Vertical Tail

In the missing vertical tail damage scenario the entire vertical tail is missing, therefore

there is a loss in directional stability and a complete loss in rudder control effectiveness
[1]. Figure 51 shows that the nominal controller response for this damage case is

unstable. Figure 52 shows that the DF-MRAC controller provides upset recovery and

satisfactory tracking performance for this damage scenario. Figure 53 shows a result

for the same damage case in Figure 52 with 0.07 seconds of time delay in the right

aileron channel. The performance of the DF-MRAC controller is not satisfactory with

this amount of time delay. Figure 54 shows that ALR modification improves the time
delay margin of the DF-MRAC controller design for this failure case.

800
800 p q
cmd 600 cmd
600 p q
400
p [deg/sec]

q [deg/sec]

400
200
200
0
0 −200
−200 −400

−400 −600

−600 −800
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time [sec] time [sec]

800 r
cmd
α
700
r
100 β
600
r [deg/sec]

α, β [deg]

500 50

400

300 0

200

100 −50

0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time [sec] time [sec]

20 1 1 1
left elev. 15
left ail. rud. left spl. stab.
0
right elev. right ail. right spl.
stabilizer [deg]

10 0.5 0.5 0.5

elevator [deg]

−5
aileron [deg]

spoiler [deg]
rudder [deg]

5
−10
0 0 0 0
−15
−5
−20
−10 −0.5 −0.5 −0.5
−25 −15

−30 −20 −1 −1 −1
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 51: GTM nominal control response for the missing vertical tail case.

89
 10
10
p q
cmd cmd
p 5
q
5

p [deg/sec]

q [deg/sec]
0 0

−5 −5

−10
−10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

4 6
r
cmd
α
5
3 r β
4
r [deg/sec]

α, β [deg]
2
3

2
1

1
0
0

−1 −1
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

6 3 1 6 0.3
left elev. left ail. rud. left spl. stab.
5 2 5
right elev. right ail. right spl. 0.2

stabilizer [deg]
0.5
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]
4
1 4
0.1
3
0 0 3
2
0
−1 2
1
−0.5
−0.1
0 −2 1

−1 −3 −1 0 −0.2
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 52: GTM DF-MRAC response for the missing vertical tail case.

600 pcmd 250 qcmd

400 p 200 q
p [deg/sec]

q [deg/sec]

200 150

100
0
50
−200
0
−400
−50
−600
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

100 r
cmd
40 α
r β
50
20
r [deg/sec]

α, β [deg]

0
0
−50
−20
−100

−150 −40

0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

20 20 1 45
3.5
left elev. 15
left ail. rud. 40 left spl. stab.
10 right elev. right ail. right spl. 3
35
stabilizer [deg]

10 0.5
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]

2.5
30
0 5
2
25
0 0 1.5
20
−10 −5 1
15
−10 −0.5 0.5
−20 10
−15 5 0

−30 −20 −1 0 −0.5

0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 53: GTM DF-MRAC response for the missing vertical tail case with 0.07
seconds of time delay in the right aileron channel.

90
 10
10
pcmd qcmd
p 5
q
5
p [deg/sec]

q [deg/sec]
0 0

−5 −5

−10
−10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

5 6
rcmd α
5
4
r β
4
r [deg/sec]

3
α, β [deg]

3
2
2
1
1

0 0

−1 −1
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time [sec] time [sec]

6 4 1 8 0.5
left elev. 3
left ail. rud. 7
left spl. stab.
5 0.4
right elev. right ail. right spl.

stabilizer [deg]
2 0.5 6
elevator [deg]

aileron [deg]

spoiler [deg]
rudder [deg]

4 0.3
1 5
3 0.2
0 0 4
2 0.1
−1 3
1 0
−2 −0.5 2
0 −3 1 −0.1

−1 −4 −1 0 −0.2
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
time [sec] time [sec] time [sec] time [sec] time [sec]

Figure 54: GTM DF-MRAC response with ALR modification term for the missing
vertical tail case with 0.07 seconds of time delay in the right aileron channel.

91
4.8 Flight Test Results on NASA AirSTAR

In this section, we analyze flight test results obtained employing a derivative-free

adaptive controller with ALR modification on the NASA Airborne Subscale Transport

Aircraft Research (AirSTAR) flight test vehicle [41, 42] shown in Figure 55, which is

designed to provide a flexible research environment for testing control algorithms

under severe uncertainties and failures. The previous section presented results on

GTM, which is a high fidelity model of AirSTAR.

In-flight damage was represented by introducing additional feedback loops to

change stability derivatives that represent unknown destabilizing changes in stiff-

ness and damping terms in the longitudinal and lateral dynamics. Here, we use the

wording adaptive controller to denote DF-MRAC with ALR modification. Figures

56–61 present samples of the results. Each figure has four subfigures. The upper

two subfigures show roll rate and pitch rate responses for a baseline, non-adaptive

control design. The lower subfigures show the roll rate and pitch rate responses for
the adaptive design. In all the figures, the dashed red line shows the pilot command

and the solid blue line shows the actual AirSTAR responses.

4.8.1 Case 1: Employing Baseline and Adaptive Controllers

Figure 56 compares the results when baseline and adaptive controllers are first en-

gaged. Here, the responses of baseline and adaptive controllers in roll and pitch rates

look nearly the same since there was no induced uncertainty, and pilot commands

occurred primarily in the pitch axis.

4.8.2 Case 2: Roll and Pitch Rate Commands

Figure 57 compares the results when a roll rate doublet and a sinusoidal pitch rate

are simultaneously commanded, without an induced uncertainty. This case shows

92
improvement in pitch rate response with adaptive control most likely due to modeling
error associated with cross axis coupling.

4.8.3 Case 3: Latency Emulation

Figure 58 compares the roll and pitch rate responses of baseline and adaptively con-

trolled AirSTAR, when a time delay up to 0.1 seconds is injected to all input channels.

It is crucial to note here that DF-MRAC with ALR modification preserves the time

delay margin of the baseline controller. In addition, if we look at the magnitude of

the commands, it can be seen that harsher commands were applied to the adaptively

controlled AirSTAR especially in roll rate. It should also be noted that the flight test
results shown in this figure were taken on different days, and that the wind conditions

for the results with DF-MRAC were considerably more variable and greater in mag-

nitude. Unfortunately we do not have a flight test result that compares DF-MRAC

with and without ALR when subjected to latency in order to validate that ALR mod-

ification does significantly improve time delay margin of an adaptive control design,

as illustrated previously in the simulated results of Figures 6 and 7.

4.8.4 Case 4: Cmα and Clp Reduction by 50% and +0.2 under Pitch Rate
Command

Figure 59 compares the roll and pitch rate responses of baseline and adaptively con-

trolled AirSTAR under parametric uncertainty (induced by introducing additional

feedback of α and p), when a pitch rate doublet is commanded. Also near the end of
the interval a pulse in the roll rate command is applied. The response with adaptive

controller is better than the baseline controller’s response. The baseline controller’s

roll rate response shown in upper left side of Figure 59 is much more oscillatory and

varies approximately from -15 to +20. The adaptive controller’s roll rate response

shown in the bottom left side of Figure 59 is better by more than a factor of 2. The

responses in pitch rate look nearly the same.

93
 Figure 55: Flight test vehicle.

4.8.5 Case 5: Cmα and Clp Reduction by 50% and +0.2 under Roll Rate
Command

Figure 60 compares the roll and pitch rate responses of baseline and adaptively
controlled AirSTAR under parametric uncertainty, when a roll rate doublet is com-

manded. Roll rate tracking with the adaptive controller is again improved by roughly

a factor of 2, and the pitch rate tracking is also improved, similar to what was observed

in Figure 57.

4.8.6 Case 6: Cmα and Clp Reduction by 75% and +0.3 under Pitch Rate
Command

Figure 61 compares the roll and pitch rate responses of baseline and adaptively con-

trolled AirSTAR under a harsher parametric uncertainty, when a pitch rate doublet

is commanded. In this case the performance in terms of tracking error in both pitch

and roll rate is reduced by at least a factor of 2 with the adaptive controller.

94
 100

15

Pitch Rate, deg/sec (Baseline)

Roll Rate, deg/sec (Baseline)
50
10

0
5

−50 0

−100 −5

660 665 670 675 680 685 690 695 660 665 670 675 680 685 690 695
Time, sec Time, sec

100

15

Pitch Rate, deg/sec (Adaptive)

Roll Rate, deg/sec (Adaptive)

50
10

0
5

−50 0

−100 −5

280 285 290 295 300 305 310 315 320 280 285 290 295 300 305 310 315 320
Time, sec Time, sec

Figure 56: Comparison of baseline and adaptive controller responses for Case 1.

4
20
Pitch Rate, deg/sec (Baseline)
Roll Rate, deg/sec (Baseline)

15 3
10
2
5
0 1
−5
0
−10
−15 −1
−20
−2
700 702 704 706 708 710 700 702 704 706 708 710
Time, sec Time, sec

4
20
Pitch Rate, deg/sec (Adaptive)
Roll Rate, deg/sec (Adaptive)

15 3
10
2
5
0 1
−5
0
−10
−15 −1
−20
−2
364 366 368 370 372 374 376 364 366 368 370 372 374 376
Time, sec Time, sec

Figure 57: Comparison of baseline and adaptive controller responses for Case 2.

95
 150 20

Pitch Rate, deg/sec (Baseline)

Roll Rate, deg/sec (Baseline)
100 15

50 10

0 5

−50 0

−100 −5

−150 −10
360 380 400 420 440 460 480 500 520 540 360 380 400 420 440 460 480 500 520 540
Time, sec Time, sec

150 20

Pitch Rate, deg/sec (Adaptive)

Roll Rate, deg/sec (Adaptive)

100 15

50 10

0 5

−50 0

−100 −5

−150 −10
400 410 420 430 440 450 460 470 480 490 400 410 420 430 440 450 460 470 480 490
Time, sec Time, sec

Figure 58: Comparison of baseline and adaptive controller responses for Case 3.

20 6
Pitch Rate, deg/sec (Baseline)
Roll Rate, deg/sec (Baseline)

4
10

2
0
0
−10
−2

−20
−4

−30 −6
784 786 788 790 792 794 784 786 788 790 792 794
Time, sec Time, sec

20 6
Pitch Rate, deg/sec (Adaptive)
Roll Rate, deg/sec (Adaptive)

4
10

2
0
0
−10
−2

−20
−4

−30 −6
408 410 412 414 416 418 408 410 412 414 416 418
Time, sec Time, sec

Figure 59: Comparison of baseline and adaptive controller responses for Case 4.

96
 30 5

Pitch Rate, deg/sec (Baseline)

Roll Rate, deg/sec (Baseline)
20

10

0 0

−10

−20

−30 −5
827 828 829 830 831 832 833 834 835 836 837 827 828 829 830 831 832 833 834 835 836 837
Time, sec Time, sec

30 5

Pitch Rate, deg/sec (Adaptive)

Roll Rate, deg/sec (Adaptive)

20

10

0 0

−10

−20

−30 −5
446 448 450 452 454 456 458 460 462 446 448 450 452 454 456 458 460 462
Time, sec Time, sec

Figure 60: Comparison of baseline and adaptive controller responses for Case 5.

30 8

6
Pitch Rate, deg/sec (Baseline)
Roll Rate, deg/sec (Baseline)

20
4
10
2

0 0

−2
−10
−4
−20
−6

−30 −8
866 868 870 872 874 876 866 868 870 872 874 876
Time, sec Time, sec

30 8

6
Pitch Rate, deg/sec (Adaptive)
Roll Rate, deg/sec (Adaptive)

20
4
10
2

0 0

−2
−10
−4
−20
−6

−30 −8
488 490 492 494 496 498 488 490 492 494 496 498
Time, sec Time, sec

Figure 61: Comparison of baseline and adaptive controller responses for Case 6.

97
4.9 Conclusion

This chapter presents a derivative-free model reference adaptive control law for un-

certain systems. The key feature is that the stability analysis is performed under
the assumption that the ideal weights are bounded, but otherwise arbitrarily time

varying. The approach is particularly useful for situations in which the nature of the

system uncertainty cannot be adequately represented by a set of basis functions with

unknown constant weights. The system error signals, including the state tracking

error and the weight update error, are proven to be uniformly ultimately bounded us-

ing a Lyapunov-Krasovskii functional without the introduction of modification terms.

We have shown that the errors approach the ultimate bound exponentially in time.

We have also provided an analysis that shows how the design parameters in the adap-

tive law can influence the ultimate bound, and the exponential rate of convergence

to the bound. The examples illustrate the superior performance of derivative free

adaptation in comparison to a conventional adaptive law, when it is evaluated for a

variety of cases in which there is a sudden or rapidly varying set of dynamics, and in
disturbance rejection. We have also shown how the derivative-free adaptive law can

be modified using many of the existing modification terms. The simulation results

show that dramatic improvement in robustness to time delays can be achieved by

modifying the derivative-free adaptation law using a method of adaptive loop recov-

ery. Furthermore, the performance of the proposed approach is successfully evaluated

in flight on AirSTAR. Finally, the proposed approach is also successfully evaluated in

flight on TwinSTAR and the results are reported in Ref. 16.

98
 CHAPTER V

Derivative-Free Adaptive Control: The Output Feedback

Case

5.1 Introduction

This chapter presents an output feedback adaptive control architecture for continuous-

time uncertain systems based on state observer and derivative-free delayed weight
update law. The observer is employed in the adaptive part of the design in place of a

reference model. A derivative-free weight update law in output feedback form ensures

that the estimated states follow both the reference model states and the true states

so that both the state estimation error and the state tracking error are bounded. The

stability analysis uses a Lyapunov-Krasovskii functional that entails the solution of

a parameter dependent Riccati equation [46], rather than a Lyapunov equation, to

show that all the error signals are UUB. The complexity of the proposed approach is

less than many other output feedback adaptive control architectures available in the

literature and it is intended to be used to augment an existing state observer based

linear controller. The proposed approach is particularly advantageous for applications

to systems that can undergo a sudden change in dynamics, such as might be due to

reconfiguration, deployment of a payload, docking, or structural damage.

99
5.2 Derivative-Free Output Feedback Adaptive Control Ar-
chitecture

Consider the uncertain system given by

 
ẋ(t) = Ax(t) + B u(t) + ∆(t, x(t)) (157)

y(t) = Cx(t) (158)

where A ∈ IRn×n , B ∈ IRn×m , and C ∈ IRm×n are known system matrices, x(t) ∈

IRn is the unknown state vector, u(t) ∈ IRm is the control input vector, ∆ : IR ×
IRn → IRm is the matched uncertainty, and y(t) ∈ IRm is the regulated output vector.

Furthermore, the triple (A, B, C) is minimal and the control input vector is restricted

to the class of admissible controls consisting of measurable functions.

Remark 5.1. The system given by (157) and (158) assumes that the control

input vector and the regulated output vector have the same dimension. For the case

when the dimension of the control input vector is larger than the dimension of the
regulated output vector and there redundancy in actuation, one can use either matrix

inverse and pseudoinverse approaches, constrained control allocation, pseudocontrols,

or daisy chaining [7, 24, 78] to reduce the dimension of the control input vector to the

dimension of the regulated output vector. Furthermore, the system can have a sensed

output vector denoted by

ys (t) = Cs x(t) (159)

where ys (t) ∈ IRl , Cs ∈ IRl×n , l ≥ m, such the elements of y(t) are a subset of the
elements of ys (t).

Consider the state observer based nominal control law given by

un (t) = −K1 x̂(t) + K2 r(t) (160)

where K1 ∈ IRm×n and K2 ∈ IRm×m are known feedback and feedforward matrices,

respectively, r(t) ∈ IRr is the regulated output command vector, and x̂(t) ∈ IRn is an

100
observer estimate of x(t) given by:

 
˙
x̂(t) = Ax̂(t) + Bun (t) + L ys (t) − Cs x̂(t) (161)

with L ∈ IRn×l being the observer gain matrix designed such that Ae ≡ A − LCs ∈
IRn×n is Hurwitz. Define the reference model as in (4) subject to Assumption 2.2.

Remark 5.2. The above implies that the gains K1 and K2 have been designed

using the certainty equivalence principle (x̂(t) = x(t)) with ∆(t, x(t)) = 0, so that

C x̂(t) tracks r(t) to within some set of specifications on both the transient and steady

state performance.

Assumption 5.1. The matched uncertainty in (157) can be linearly parameter-

ized as

∆(t, x) = W T (t)β(x), ∀x ∈ Dx ⊂ IRn (162)

where W (t) ∈ IRs×m is the unknown time-varying weight matrix that satisfies kW (t)k ≤

ω̄, β : IRn → IRs is a known Lipschitz continuous basis vector of the form β(x) =

[β1 (x), β2 (x), ..., βs (x)]T with |β(x)| ≤ β̄, and Dx is sufficiently large.

Remark 5.3. Assumption 5.1 expands the class of uncertainties that can be
represented by a given set of basis functions. That is, an adaptive law design subject

to Assumption 5.1 can be more effective than an adaptive law designed subject to

∆(x) = W T β(x) + ε(x), ∀x ∈ Dx (163)

where W is a unknown constant ideal weight and ε(x) is the residual error. It also

permits an explicit dependence of the uncertainty on time.

Remark 5.4. Assumption 5.1 does not place any restriction on the time derivative

of the weight matrix. However the degree of time dependence will depend on how
β(x) is chosen.

The adaptive control objective is to design a control law u(·) for the system given

by (157) and (158) so that x(t) tracks xm (t) with bounded error. For this purpose,

101
the nominal control law un (t) given by (160) is augmented with the adaptive control
law uad (t) as:

u(t) = un (t) − uad (t) (164)

uad (t) = Ŵ T (t)β(x̂(t)) (165)

Note that the state observer given by (161) is regarded as a part of the nominal

control design. However, our viewpoint below is that L may be altered for purposes

of adaptively augmenting the nominal controller.

Denote x̃(t) ≡ x(t) − x̂(t) for the state estimation error, ê(t) ≡ x̂(t) − xm (t) for the
estimated state tracking error, and W̃ (t) ≡ W (t)− Ŵ (t) for the weight estimate error.

From (157) and (161), the dynamics for the state estimation error can be written in

the form

˙
x̃(t) = Ae x̃(t) + B W̃ T (t)β(x̂(t)) + Bg(x(t), x̂(t)) (166)

 
where g(x(t), x̂(t)) ≡ W T (t) β(x̂(t))−β(x(t)) with |g(x(t), x̂(t))| ≤ w̄Lβ |x̃(t)|, where

Lβ > 0 is the Lipschitz constant for the known basis vector. Likewise, from (161) and

(4) the dynamics for the estimated state tracking error can be written in the form:

˙
ê(t) = Am ê(t) + LC x̃(t) (167)

Next, consider the derivative-free weight update law given by

Ŵ (t) = Ω1 Ŵ (t − τ ) + Ω̂2 (t) (168)

where τ > 0, and Ω1 ∈ IRs×s and Ω̂2 : IRn × IRn → IRs×m satisfy:

0 ≤ ΩT
1 Ω1 < κ1 I, 0 ≤ κ1 < 1/(1 + µ), µ>0 (169)

Ω̂2 (t) ≡ κ2 β(x̂(t))ỹ T (t), κ2 > 0 (170)

where ỹ(t) ≡ y(t) − ŷ(t), where ŷ(t) = C x̂(t).

102
 Figure 62: Derivative-free output feedback adaptive control architecture

A visualization of the derivative-free output feedback adaptive control architecture

is given in Figure 62. Note that the state observer serves as the reference model. Its

dynamics are the same as the reference model if uad (t) cancels ∆(t, x(t)), and in this

case the observer error transient ỹ(t) goes to zero. The components that are added

to an observer based nominal controller architecture, in order to realize the adaptive

control, reduce to computing the basis functions and implementing the adaptation
law in (168).

5.3 A Parameter Dependent Riccati Equation

In this section we summarize the properties of a parameter dependent Riccati equation

[46] that is needed for the stability analysis of the following section. Consider the

following quadratic equation in P

0 = AT
e P + P Ae + Q (171)

Q = Q0 + (κ2 + 1/µ)β̄ 2NN T = Q0 + νNN T (172)

N = CT − P B (173)

103
in which Q0 ∈ IRn×n > 0, and µ and β̄ have previously been defined by (169) and
Assumption 5.1, respectively.

Remark 5.5. Let 0 < ν < ν̄ define the largest set within which there exists a

positive definite solution for P . Since P > 0 for ν = 0 and P depends continuously

on ν, the existence of P (ν) > 0 for 0 < ν < ν̄ is assured. Furthermore, this implies

that κ2 < ν̄/β̄ 2 − 1/µ ≡ κ̄2 .

Remark 5.6. If N = 0 in (173), then it follows from (171)–(173) that we have

0 = AT
e P + P Ae + Q0 (174)

0 = CT − P B (175)

which implies that the transfer function associated with the system G(s) = C(sI −

Ae )−1 B is positive-real [30]. In this case (171) reduces to a Lyapunov equation asso-

ciated with the error dynamics in (166), which is commonly employed in the stability

analysis of adaptive systems.

Lemma 5.1. If Am has no repeated eigenvalues and the state observer gain matrix
L of (161) is designed using pole placement such that

λ(Ae ) = kλ(Am ) (176)

then for ν < ν̄, we can denote P (k) as the corresponding positive definite solution for

P of the parameter dependent Riccati equation given by (171)–(173). In this case:

lim P (k) = 0 (177)

k→∞

Proof. Let Sm and Se be diagonalizing transformations for Am and Ae , respec-

tively. Then:

−1
Am = Sm diag[λ(Am )]Sm (178)

Ae = kSe diag[λ(Am )]Se−1 (179)

104
 −1
With Ss ≡ Se Sm , Ae can be written as:

Ae = kSs Am Ss−1 = kAs (180)

Using (180) and denoting Pk ≡ kP (k) and Nk ≡ C T − k −1 Pk B, the parameter

dependent Riccati equation in (171)–(173) can be written as:

0 = AT T
s Pk + Pk As + Q0 + νNk Nk (181)

Taking the limit of (181) as k → ∞, it follows that:

0 = AT
s P∞ + P∞ As + Q0 + νCC
T
(182)

Since the solution of (182) is finite, it follows that for all finite k, P (k) > 0, and:

Pk
lim P (k) = lim =0 (183)
k→∞ k→∞ k

Remark 5.7. Lemma 5.1 implies that ||P B|| can be made arbitrarily small by

making the observer dynamics sufficiently fast, while preserving the property that
P > 0. This property will become important in the stability analysis of Section IV.

The next lemma shows that for ν < ν̄, (171)–(173) can reliably be solved for

P > 0 using the Potter approach given in Ref. 67. This also implies that ν̄ can

be determined by searching for the boundary value that results in a failure of the

algorithm to converge. We employ the notation ric(·) and dom(ric) as defined in Ref.

22.
Lemma 5.2. Let P satisfy the parameter dependent Riccati equation given by

(171)–(173) and let the modified Hamiltonian be given by

 
Ae − νBC νR 
H≡
T  (184)
−Q − Ae − νBC

where Q ≡ Q0 + νCC T and R ≡ BB T . Then, for all 0 < ν < ν̄, H ∈ dom(ric) and

P = ric(H).

105
 Proof: The proof follows from Lemmas 1 and 2 of Ref. 22. 

5.4 Stability Analysis

This section presents a stability analysis for the derivative-free output feedback adap-

tive control architecture.

Theorem 5.1. Consider the controlled uncertain dynamical system given by

(157) and (158) subject to Assumption 5.1. Consider, in addition, the feedback
control law given by (164), with the nominal feedback control component given by

(160) and (161), and the adaptive feedback control component given by (165) and

(168) subject to the conditions in (169) and (170), and with κ2 < κ̄2 . Then, x̃(t) and

W̃ (t) are UUB.

Proof. Using (168) and defining

Ω2 (t) ≡ W (t) − Ω1 W (t − τ ) (185)

where ||Ω2 (t)|| ≤ δ̄, δ̄ ≡ w̄(1 + ||Ω1 ||), the weight update error W̃ (t) can be rewritten
as:

W̃ (t) = Ω1 W̃ (t − τ ) + Ω2 (t) − Ω̂2 (t) (186)

Using (186), the state estimation error in (166) becomes:

 T
˙
x̃(t) = Ae x̃(t) + B Ω1 W̃ (t − τ ) + Ω2 (t) − Ω̂2 (t) β(x̂(t))

+Bg(x(t), x̂(t)) (187)

Note that Ω̂2 (t), given by (170), can be equivalently written using (173) as:

Ω̂2 (t) = κ2 β(x̂(t))x̃T (t)C T

 
= κ2 β(x̂(t))x̃T (t) P B + N

= κ2 β(x̂(t))x̃T (t)P B + κ2 β(x̂(t))x̃T (t)N (188)

106
 To show that the closed-loop system given by (186) and (187) is UUB, consider
the Lyapunov-Krasovskii functional [31]
Z t
T
 
V(x̃(t), W̃t ) = x̃ (t)P x̃(t) + ρ tr W̃ T (s)W̃ (s)ds (189)
t−τ

where ρ > 0, W̃t represents W̃ (t) over the time interval t−τ to t, and P satisfies (171)

with ν < ν̄, which implies κ2 < κ̄2 as noted in Remark 5.5. The directional derivative

of (189) along the closed-loop system trajectories of (186) and (187) is given by

V̇(x̃(t), W̃t ) = −x̃T (t)Qx̃(t) + 2x̃T (t)P Bg(x(t), x̂(t))

+2x̃T (t)P B[Ω1 W̃ (t − τ )]T β(x̂(t))

−2x̃T (t)P B Ω̂T T T

2 (t)β(x̂(t)) + 2x̃ (t)P BΩ2 β(x̂(t))
 
+ρ tr −ξ W̃ T (t)W̃ (t) + η W̃ T (t)W̃ (t) − W̃ T (t − τ )W̃ (t − τ ) (190)

where η ≡ 1 + ξ, ξ > 0. Using (186) to expand the term tr[η W̃ T (t)W̃ (t)] in (190)

produces:

V̇(x̃(t), W̃t ) = −x̃T (t)Qx̃(t) + 2x̃T (t)P Bg(x(t), x̂(t))

+2x̃T (t)P B[Ω1 W̃ (t − τ )]T β(x̂(t))

a
−2x̃T (t)P B Ω̂T
2 (t)β(x̂(t)) + 2x̃T (t)P BΩT 2 β(x̂(t))

+ρ tr −ξ W̃ T (t)W̃ (t) − W̃ T (t − τ )W̃ (t − τ )

+η W̃ T (t − τ )ΩT
1 Ω1 W̃ (t − τ )
a a
+η Ω̂T
2 (t)Ω̂2 (t) + ηΩT T
2 Ω2 −2η Ω̂2 (t)Ω1 W̃ (t − τ )
b a
+2η W̃ T (t − τ )ΩT Ω
1 2 −2η Ω̂T
2 (t)Ω2 (191)

Consider |aT b| ≤ γaT a + bT b/4γ, γ > 0, that follows from Young’s inequality [5]

extended to the vector case for any vectors a and b. This can be further gener-
 
alized to |tr AT B | = |vec(A)T vec(B)| ≤ γvec(A)T vec(A) + vec(B)T vec(B)/4γ =

107
    
γtr AT A +tr B T B /4γ, γ > 0, for any matrices A and B with appropriate dimen-
sions. Using this, we can write:

   
tr 2η W̃ T (t − τ )ΩT
1 Ω2 (t) ≤ tr γ W̃ T (t − τ )ΩT 1 Ω1 W̃ (t − τ )
 
+tr η 2 ΩT 2 (t)Ω2 (t)/γ , γ>0 (192)

a b
Using (188) for the terms in · and (192) for the term in · , (191) becomes:

V̇(x̃(t), W̃t ) ≤ −x̃T (t)Qx̃(t) + 2x̃T (t)P Bg(x(t), x̂(t))

c
+2x̃T (t)P B[Ω1 W̃ (t − τ )]T β(x̂(t))
f
−2x̃T (t)P B[κ2 β(x̂(t))x̃T (t)P B]T β(x̂(t))
d
−2x̃T (t)P B[κ2 β(x̂(t))x̃T (t)N]T β(x̂(t))
e
+2x̃T (t)P BΩT2 β(x̂(t))
 
+ρ tr −ξ W̃ T (t)W̃ (t)
  g
+ρ tr −W̃ T (t − τ )W̃ (t − τ )
  g
+ρ tr η W̃ T (t − τ )ΩT1 Ω1 W̃ (t − τ )
  f
+ρ tr ηκ22 B T P x̃(t)β T (x̂(t))β(x̂(t))x̃T (t)P B
  d̄
+2ρ tr ηκ22 B T P x̃(t)β T (x̂(t))β(x̂(t))x̃T (t)N
 
+ρ tr ηκ22 N T x̃(t)β T (x̂)β(x̂)x̃T (t)N
  h
+ρ tr ηΩT 2 Ω 2
h  T i c̄
+ρ tr −2η κ2 β(x̂(t))x̃T (t)P B Ω1 W̃ (t − τ )
h  T i
T
+ρ tr −2η κ2 β(x̂(t))x̃ (t)N Ω1 W̃ (t − τ )
  g
+ρ tr γ W̃ T (t − τ )ΩT 1 Ω1 W̃ (t − τ )
  h
+ρ tr η 2 ΩT Ω
2 2 /γ
h  T
T i ē
+ρ tr −2η κ2 β(x̂(t))x̃ (t)P B Ω2
h  T i
T
+ρ tr −2η κ2 β(x̂(t))x̃ (t)N Ω2 (193)

108
 c
Since κ2 = 1/ρη, · cancels · c̄ , · d
cancels · d̄
, and · e
cancels · ē
in the above

expression. Furthermore, grouping the terms in · f as −κ2 x̃T (t)P BB T P x̃(t) β T (x̂(t))
 
β(x̂(t)), the terms in · g as −ρ tr W̃ T (t − τ )[I − (η + γ)ΩT
1 Ω1 ]W̃ (t − τ ) , and the
h
terms in · as ρ(η + η 2 /γ) tr[ΩT
2 Ω2 ] yields

V̇(x̃(t), W̃t ) ≤ −x̃T (t)Qx̃(t) + 2x̃T (t)P Bg(x(t), x̂(t))

−κ2 x̃T (t)P BB T P x̃β T (x̂(t))β(x̂(t))

 
−ρξ tr[W̃ T (t)W̃ (t)] − ρ tr W̃ T (t − τ )[I − (η + γ)ΩT
1 Ω1 ]W̃ (t − τ )

+ρ(η + η 2 /γ) tr[ΩT T T

2 Ω2 ] − ρκ2 tr[2ηN x̃(t)β (x̂(t))Ω2 ]

+κ2 x̃T (t)NN T x̃(t)β̄ 2 − 2 tr[N T x̃(t)β T (x̂(t))Ω1 W̃ (t − τ )] (194)

where using Young’s inequality again for the last term in (194) produces:

2 tr[N T x̃(t)β T (x̂(t))Ω1 W̃ (t − τ )] ≤ tr[µW̃ T (t − τ )ΩT

1 Ω1 W̃ (t − τ )]

+x̃T (t)NN T x̃(t)β̄ 2 /µ, µ>0 (195)

Using (195) in (194) results in

V̇(x̃(t), W̃t ) ≤ −x̃T (t)Qx̃(t) + 2x̃T (t)P Bg(x(t), x̂(t))

−κ2 x̃T (t)P BB T P x̃β T (x̂(t))β(x̂(t)) − ρξ tr[W̃ T (t)W̃ (t)]

 
−ρ tr W̃ T (t − τ )[I − (η + γ + µ)ΩT 1 Ω1 ]W̃ (t − τ )

+ρ(η + η 2 /γ) tr[ΩT T T

2 Ω2 ] − ρκ2 tr[2ηN x̃(t)β (x̂(t))Ω2 ]


κ2 + 1/µ β̄ 2 x̃T (t)NN T x̃(t) (196)

Applying the definitions of Q and ν in (172), and enforcing κ2 < κ̄2 , the first and last
terms in (196) can be combined to produce the following inequality

V̇(x̃(t), W̃t ) ≤ −x̃T (t)Q0 x̃(t) + 2x̃T (t)P Bg(x(t), x̂(t))

−κ2 x̃T (t)P BB T P x̃β T (x̂(t))β(x̂(t)) − ρξ tr[W̃ T (t)W̃ (t)]

 
−ρ tr W̃ T (t − τ )[I − (η + γ + µ)ΩT 1 Ω1 ]W̃ (t − τ )

+ρ(η + η 2 /γ) tr[ΩT T T

2 Ω2 ] − ρκ2 tr[2ηN x̃(t)β (x̂(t))Ω2 ] (197)

109
  
Define κ1 ≡ 1/(η+γ +µ). Note that tr W̃ T (t−τ )[I −κ−1 T
1 Ω1 Ω1 ]W̃ (t−τ ) = vec W̃ (t−

T

T


τ ) vec [I − κ−1 T
1 Ω1 Ω1 ]W̃ (t − τ ) = vec W̃ (t − τ ) I ⊗ [I − κ−1 T
1 Ω1 Ω1 ] vec W̃ (t − τ )


≥ λmin I ⊗ [I − κ−1 T 2 −1 T 2
1 Ω1 Ω1 ] ||W̃ (t − τ )|| = λmin (I − κ1 Ω1 Ω1 )||W̃ (t − τ )|| . Using this

and (172) in (197) with |g(x(t), x̂(t))| ≤ ḡ, ḡ ≡ 2w̄β̄, yields

V̇(x̃(t), W̃t ) ≤ −c1 |x̃(t)|2 − c2 ||W̃ (t)||2 − c3 ||W̃ (t − τ )||2 + d1 + d2 |x̃(t)| (198)

where c1 ≡ λmin (Q0 ) > 0, c2 ≡ ρξ > 0, c3 ≡ ρλmin (I − κ−1 T

1 Ω1 Ω1 ) > 0, d1 ≡

ρ(η + η 2 /γ)δ̄ 2 ≥ 0, and d2 ≡ 2(||P B||ḡ + ||N||||Ω2||β̄) > 0. We can further arrange

(198) as

√ √ 2
V̇(x̃(t), W̃t ) ≤ − c1 |x̃(t)| − d2 /2 c1 −c2 ||W̃ (t)||2 − c3 ||W̃ (t − τ )||2 + d (199)

where d ≡ d1 +d22 /4c1 . Either |x̃(t)| > Ψ1 or kW̃ (t)k > Ψ2 or kW̃ (t−τ )k > Ψ3 renders
p p p
V̇(x̃(t), W̃t ) < 0, where Ψ1 ≡ d/c1 + d2 /2c1 , Ψ2 ≡ d/c2 , and Ψ3 ≡ d/c3. Hence,

it follows that x̃(t) and W̃ (t) are UUB. 

Remark 5.8. The derivative-free weight update law given by (168) does not

require a modification term to prove the error dynamics, including the weight errors,

are UUB.

Remark 5.9. Derivative-free adaptive control does not employ an integrator in

its weight update law. This is advantageous from the perspective of augmenting a

nominal controller that employs integral action to ensure that the regulated output

variables track r(t) for constant disturbances, regardless of how these disturbances

may enter the system. An example that illustrates this advantage is provided for

full-state feedback case in Section V of Ref. 85.

R 0 
Define q(t) ≡ [x̃T (t), ṽ(t, τ )]T , where ṽ 2 (t, τ ) ≡ tr −τ
W̃ T (s)W̃ (s)ds , and let
Br = {q(t) : |q(t)| < rα }, such that Br ⊂ Dq for a sufficiently large compact set Dq .

Then, we have the following corollary.

Corollary 5.1. Under the conditions of Theorem 5.1, an estimate for the ultimate

110
bound for q(t) is given by
s
λmax (P )Ψ21 + ρτ Ψ22
r= (200)
λmin (P̃ )

where P̃ = diag[P, ρ].

min
Proof. Denote Ωα = {q(t) ∈ Br : q T (t)P̃ q(t) ≤ α̂}, α̂ = ||q(t)||=r
q T (t)P̃ q(t) =

rα2 λmin (P̃ ). Since

V(x̃(t), W̃t ) = q T (t)P̃ q(t)

Z t
T
 
= x̃ (t)P x̃(t) + ρ tr W̃ T (s)W̃ (s)ds (201)
t−τ

it follows that Ωα is an invariant set if and only if:

α̂ ≥ λmax (P )Ψ21 + ρτ Ψ22 (202)

Thus, the minimum size of Br that ensures this condition has radius given by (200).


Remark 5.10. The proofs of Theorem 5.1 and Corollary 5.1 assume that the sets

Dx and Dq are sufficiently large. If we define Br∗ as the largest ball contained in Dq ,

and assume that the initial conditions are such that q(0) ⊂ Br∗ , then from Figure 2

we have added the condition that r < r ∗ , which implies an lower bound on ρ. It can

be shown that in this case the lower bound must be such that λmin (P̃ ) = ρ. With r
defined by (200) and λmin (P̃ ) = ρ, the condition r < r ∗ implies:

λmax (P )Ψ21
ρ > (203)
r ∗2 − τ Ψ22

Since κ2 = 1/ρη, η > 1, it follows from (203) that r ∗ should ensure that:

r ∗2 − τ Ψ22
κ2 < (204)
λmax (P )Ψ21

Therefore, the meaning of Dq sufficiently large in Corollary 5.1 is that

q
r∗ > κ2 λmax (P )Ψ21 + τ Ψ22 (205)

111
and q(0) ⊂ Dr∗ . The meaning of Dx sufficiently large is difficult to characterize
precisely since x(t) depends on both r(t) and x(0). Nevertheless it can be seen that

increasing κ2 implies increasing the require size of the set Dx .

Lemma 5.3. If x̃(t) is bounded, then the state tracking error defined as e(t) ≡

x(t) − xm (t) is bounded.

Proof.

|e(t)| = |x(t) − xm (t)|

= |x(t) − x̂(t) + x̂(t) − xm (t)|

≤ |x(t) − x̂(t)| + |x̂(t) − xm (t)|

= |x̃(t)| + |ê(t)| (206)

where if x̃(t) is bounded then from (167) ê(t) is bounded. This implies that e(t) is

bounded. 
Corollary 5.2. Consider the system of equations given by (166) and (167). If

x̃(t) is UUB by r, then e(t) is UUB by r(1 + v) where

2||Pm LC||
v ≡ (207)
λmin(Qm )
and Pm ∈ IRn×n being a positive-definite solution to the Lyapunov equation given by

0 = AT
m Pm + Pm Am + Qm (208)

for a given positive-definite matrix Qm ∈ IRn×n .

Proof. To show that ê(t) is UUB, consider the Lyapunov function given by

V(ê(t)) = êT (t)Pm ê(t) (209)

where Pm satisfies (208). We can write the directional derivative of (209) along the

trajectories of ê(t) given by (167) as

V̇(ê(t)) = −êT (t)Qm ê(t) + 2êT (t)Pm LC x̃(t)

h i
≤ −|ê(t)| λmin (Qm )|ê(t)| − 2||Pm LC|| |x̃(t)|

112
so that V̇(ê(t)) < 0 for all |ê(t)| > rv, since x̃(t) is UUB by r and v is given by (207).
Hence, it follows from Lemma 5.3 that |e(t)| is UUB by r(1 + v). 

Remark 5.11. It follows from Corollaries 4.1 and 4.2 that |e(t)| is UUB by

re1 ≡ r(1 + v), where:

s " #
λmax (P )Ψ21 + ρτ Ψ22 2||Pm LC||
re 1 = 1+ (210)
λmin (P̃ ) λmin (Qm )

5.5 Examples on a Model of Wing Rock Dynamics

This section gives several examples of the architecture described in Figure 62 using

a model of wing rock dynamics. A two state model for wing rock dynamics can be

written in the form given by (32) and

 
 
x1 (t)
ys (t) = 1 0   + w(t) (211)
x2 (t)

where w(t) is measurement noise disturbance and

   
∆(t, x(t)) = α1 + f1 (t) x1 (t) + α2 + f2 (t) x2 (t) + α3 |x1 (t)|x2 (t)

+α4 |x2 (t)|x2 (t) + α5 x31 (t) + d(t) (212)

with constant coefficients α1 = 0.2314, α2 = 0.6918, α3 = −0.6245, α4 = 0.0095,

and α5 = 0.0214, and time varying coefficients f1 (t) and f2 (t), and an external

disturbance d(t). We chose K1 = [2.56, 2.56] and K2 = 2.56 for the nominal controller

design, and LT = [12.8 64.0] for the observer gain which corresponds to k = 5 in

(176). For the adaptive control design, we used sigmoidal basis functions of the form
  1−e−xi (t)
β(x(t)) = 0.5, β1 (x1 (t)), β2 (x2 (t)) , where βi (xi (t)) = 1+e −xi (t) , i = 1, 2. Since

|βi (xi )| ≤ 1, it follows that β̄ = 1.5. For µ = 0.05 and Q0 = 0.25I2 , it was determined

by Lemma 5.2 that ν̄ = 124.6 which implies that the adaptation gain in (170) must

satisfy κ2 < 35.4. Furthermore, we set Ω1 = 0.95I3 , κ2 = 35, and τ = 0.01 seconds.

113
 We consider the command tracking problem in the following subsections with the
initial conditions for the dynamics in (32) set zero. Figure 63 shows the nominal and

adaptive control responses for the case when the ideal weights are constant f1 (t) =



f2 (t) = 0 , and in the absence of disturbances d(t) = w(t) = 0 . Overall performance

is significantly improved with adaptation. Improvement is measured by comparing

how well the responses follow ym (t). Figure 64 compares the responses for the case

when the ideal weights are time varying, with f1 (t) being a square wave having an
amplitude of 0.5 and a period of 15 seconds and f2 (t) = 0.5sin(1.5t). The improvement

with adaptation is greater in comparison to the improvement in Figure 63. The effect

of f1 (t) is clearly evident in the nominal case.

Next we include the effect of a process disturbance and measurement noise. In this

case d(t) is a square wave having an amplitude of 0.1 and a period of 6 seconds. To

model sensor noise we let w(t) be a band-limited white noise process with a correlation
time constant of 0.01 seconds, and a noise power level of 0.0001. These processes are

depicted in Figure 65. Figure 66 shows that tracking performance is significantly

improved with adaptation, and that the control time history is well behaved (sensor

noise is not amplified). Figure 67 includes a comparison between ∆(t, x(t)) and uad (t).

5.6 Examples on an Aeroelastic Generic Transport Model

This section illustrates an application of the proposed approach on an aeroelastic

model of longitudinal dynamics for a generic transport model. The form of the

derivative-free, output feedback adaptive control differs from the one given in Section
5.2 in that it is designed to augment a nominal controller with integral action. For

this purpose, we consider the uncertain system given by

 
ẋp (t) = Ap xp (t) + Bp u(t) + ∆(t, xp (t)) (213)

y(t) = Cp xp (t) (214)

114
 6 ym(t)
y(t), Adaptation On
4
y(t), Adaptation Off

ym (t), y(t)
2

−2

−4

−6

0 5 10 15 20 25 30
Time, s

0.4
0.3
0.2
0.1
u(t)

0
−0.1
−0.2
−0.3
−0.4

0 5 10 15 20 25 30
Time, s

Figure 63: Nominal and adaptive control responses for the case of constant ideal
weights

10
ym(t)
y(t), Adaptation On
y(t), Adaptation Off
ym (t), y(t)

−5

0 5 10 15 20 25 30
Time, s

0.4

0.2

0
u(t)

−0.2

−0.4

−0.6
0 5 10 15 20 25 30
Time, s

Figure 64: Nominal and adaptive control responses for the case of time varying ideal
weights

115
 0.1

0.05

d(t)
0

−0.05

−0.1
0 5 10 15 20 25 30
Time, s

0.3

0.2

0.1
w(t)

−0.1

−0.2

−0.3

0 5 10 15 20 25 30
Time, s

Figure 65: Depiction of d(t) and w(t)

15 y (t)
m
y(t), Adaptation On
10 y(t), Adaptation Off
ym (t), y(t)

−5

−10

0 5 10 15 20 25 30
Time, s

0.5
u(t)

−0.5
0 5 10 15 20 25 30
Time, s

Figure 66: Nominal and adaptive control responses with disturbances for the case of
time varying ideal weights

116
 0.3
Delta(t,x(t))

∆(t, x(t)), uad (t)

0.2 u (t)
ad

0.1

−0.1

−0.2

0 5 10 15 20 25 30
Time, s

Figure 67: ∆(t, x(t)) and uad (t)

where Ap ∈ IRnp ×np , Bp ∈ IRnp ×m , and Cp ∈ IRm×np are known system matrices with
(Ap , Bp , Cp) being the minimal triple, xp (t) ∈ IRnp is the unknown state vector, u(t) ∈

IRm is the control input vector restricted to the class of admissible controls consisting

of measurable functions, ∆ : IR × IRnp → IRm is the matched uncertainty satisfying

Assumption 5.1, and y(t) ∈ IRm is the regulated output vector. Furthermore, the

system can have a sensed output vector denoted by

ys (t) = Cs xp (t) (215)

where ys (t) ∈ IRlp , Cs ∈ IRlp ×np , lp ≥ m, such the elements of y(t) are a subset of the

elements of ys (t).
We will consider the situation in which there is a state observer based nominal

controller in which the control of the regulated outputs that are commanded include

integral action, whereas the regulated outputs that are not commanded are subject

to proportional control. Let

   
y1 (t) Cp1 
y(t) =   =  xp (t) (216)
y2 (t) Cp2

where y1 (t) ∈ IRr , r ≤ m, is regulated with proportional and integral control to track

a given command vector r(t) ∈ IRr , y2 (t) ∈ IRm−r is regulated with proportional

control, Cp1 ∈ IRr×np , and Cp2 ∈ IR(m−r)×np . The integrator dynamics are defined by

ẋint (t) = −y1 (t) + r(t) = −Cp1 xp (t) + Ir r(t) (217)

117
where xint (t) ∈ IRr . Considering (213), (214), and (217), the augmented dynamics
become:
     
 Ap 0 Bp  0 
h i
ẋ(t) =   x(t) +   u(t) + ∆(t, xp (t)) +   r(t) (218)
−Cp1 0 0 I
| {z } | {z } |{z}
A B Bm
 
y(t) = Cp 0 x(t) (219)
| {z }
C

where x(t) = [x(t), xint (t)]T ∈ IRn , A ∈ IRn×n , B ∈ IRn×m , Bm ∈ IRn×r , C ∈ IRm×n ,

and n = np + r. In addition, the augmented sensed output vector becomes

   
 ys (t)   Cs 0 
ȳs (t) =   =   x(t) (220)
xint (t) 0 Ir
| {z }
C̄s

where ȳs (t) ∈ IRl , C̄s ∈ IRl×n , l = lp + r. Consider the state observer based nominal

control law given by

un (t) = −K x̂(t) (221)

where K ∈ IRm×n is known feedback matrix and x̂(t) ∈ IRn is an observer estimate
of x(t) given by

 
˙
x̂(t) = Ax̂(t) + Bun (t) + Bm r(t) + L ȳs (t) − C̄s x̂(t) (222)

with L ∈ IRn×l being the observer gain matrix designed such that Ae ≡ A − LC̄s ∈

IRn×n is Hurwitz.

Define the reference model as in (4) subject to Am = A − BK being Hurwitz by

design. This implies that the gain K has been designed using the certainty equiva-

lence principle (x̂(t) = x(t)) with ∆(t, xp (t)) = 0, so that y(t) is regulated and y1 (t)

tracks r(t) to within some set of specifications on both the transient and steady state

performance.

118
 The adaptive control objective is to design a control law u(·) for the system given
by (218) and (219) so that x(t) tracks xm (t). For this purpose, the nominal control

law un (t) given by (221) is augmented with the adaptive control law uad (t) as in (164)

with

uad (t) = Ŵ T (t)β(x̂p (t)) (223)

 
where x̂p (t) = Inp , 0np ×m x̂(t).

Consider the derivative-free weight update law given by (168), where τ > 0, and

Ω1 ∈ IRs×s and Ω̂2 : IRn × IRn → IRs×m satisfy (169) and

Ω̂2 (t) ≡ κ2 β(x̂p (t))ỹ T (t), κ2 > 0 (224)

where ỹ(t) ≡ y(t) − ŷ(t), where ŷ(t) = C x̂(t).

A visualization of the derivative-free output feedback adaptive control architecture

is given in Figure 68. Its dynamics are the same as the reference model if uad (t) cancels

∆(t, xp (t)), and in the case the observer error transient ỹ(t) goes to zero.

Denote x̃(t) ≡ x(t) − x̂(t) for the state estimation error, ê(t) ≡ x̂(t) − xm (t) for the
estimated state tracking error, and W̃ (t) ≡ W (t)− Ŵ (t) for the weight estimate error.

From (218) and (222), the dynamics for the state estimation error can be written in

the form

˙
x̃(t) = Ae x̃(t) + B W̃ T (t)β(x̂(t)) + Bg(xp (t), x̂p (t)) (225)
 
where g(xp (t), x̂p (t)) ≡ W T (t) β(x̂p (t))−β(xp (t)) with |g(xp(t), x̂p (t))| ≤ w̄Lβ |x̂p (t)−

xp (t)| ≤ w̄Lβ |x̃(t)|, where Lβ > 0 is the Lipschitz constant for the known basis vector.

Likewise, from (222) and (4) the dynamics for the estimated state tracking error can
be written in the form:

˙
ê(t) = Am ê(t) + LC x̃(t) (226)

Hence, it follows identical to the proof of Theorem 5.1 that x̃(t) and W̃ (t) are UUB,

and then to the proof of Corollary 5.2 that e(t) is UUB.

119
 Figure 68: Derivative-free output feedback adaptive control architecture

We can now formulate a state space model of the rigid body pitch dynamics

coupled with the flexible body dynamics. The details for this modeling are given in

Ref. 59. For the configuration with 50% fuel, and with an altitude of 30000 feet and


a Mach number of 0.8, a linearized model under nominal conditions ∆(t, xp (t)) = 0

120
is obtained in the form of (213) and (214) with
 
 −8.01 × 10 −1
9.65 × 10 −1
1.26 × 10 −2
5.09 × 10 −1
5.46 × 10 −4
−2.42 × 10 −3

 
 
 

 −2.45 −9.14 × 10−1 1.75 × 10−1 7.39 9.11 × 10−3 −3.11 × 10−2 

 
 
0 0 0 0 1 0
 
 
Ap =   (227)
 
 

 0 0 0 0 0 1 

 
 
 1.42 × 103 3.96 × 101 −3.13 × 101 −1.40 × 103
 
 −3.25 −8.26 

 
 
−2.62 × 102 −5.63 3.78 −1.89 × 102 2.52 × 10−1 −3.75

" #T
Bp = 6.50 × 10−2 3.52 0 0 0 0 (228)

 
 0 1 0 0 0 0 
Cp =  (229)
 

 
1 1
−1.98 × 10 −8.47 × 10 −1
3.11 × 10 −1
1.26 × 10 1.35 × 10 −2
−6.00 × 10 −2

with the state vector being defined as xp (t) = [α(t), q(t), w(t), θ(t), ẇ(t), θ̇(t)]T ,

where α(t) denotes the angle of attack, q(t) denotes the pitch angular rate, w(t)

denotes the bending modal amplitude, and θ(t) denotes the torsional modal am-

plitude. The control input and the sensed output are defined as u(t) = δe (t) and

ys (t) = [q(t), Az (t)]T , where δe (t) denotes the elevator deflection and Az (t) denotes

the normalized acceleration at the aircraft center of gravity. The measurement equa-
tion for Az (t) was obtained using the relationship

Az (t) = U0 α̇(t) − q(t) /g (230)

where U0 is the equilibrium speed and g is the acceleration due to gravity. Only q(t)

is regulated, so the dimension of y(t) = y1 (t) in (216) and of r(t) in (4) is one (r = 1).

Therefore, the dimension of the augmented sensed output vector, ȳs (t) in (220), is

three (l = 3). Note that since the second entry of Bp in (228) is approximately

50 times greater than its first entry, the uncertainty ∆(t, xp (t)) introduced in the
following subsections is approximately matched.

121
 We included the effect of actuator dynamics, measurement noise, and analog pre-
filtering of the ys (t) in our simulation. The actuator model for the elevator has a

bandwidth of 12Hz, an amplitude saturation of ±30 degrees, and a rate saturation

of ±100 degrees per second. To model sensor noise, we assumed independent band-

limited white noise processes with correlation time constants of 0.001 seconds and

noise power levels of 1.5 × 10−8 and 1 × 10−6 , respectively. The pre-filters for q(t)

and Az (t) each have a bandwidth of 8Hz, which is well beyond the bandwidth of the
nominal control design.

The eigenvalues of the rigid aircraft’s short period mode, obtained from the 2 × 2

upper left matrix partition of (227), are −0.85±1.53. The eigenvalues of the aircraft’s

two flexible modes, obtained from the 4 × 4 lower right matrix partition of (227), are

−3.32 ± 7.56 and −0.18 ± 12.83. These correspond to a bending and a torsion mode,

respectively. The spectrum of Ap is ρ(Ap ) = {−4.53 ± 8.07, 0.71 ± 13.08,−2.63, and

1.56}. This shows that the linearized model under nominal conditions is unstable due

to coupling effects and the low damping of the torsion mode. The frequency responses

from the control input to the two sensed outputs for this model are given in Figure

69. The spike in the bode plot of the transfer function from δe (t) to q(t) is due to a

near pole/zero cancellation associated with the torsion mode. The resonance due to

the torsion mode is evident in the transfer function from δe (t) to Az (t).
LQG/LTR theory was used to design the nominal controller [23]. The controller

gain matrix (K) was obtained using QK = diag[1, 25, 0.0001, 5, 0.0001, 5, 125] to pe-

nalize x(t), and RK = 10 to penalize u(t). The observer gain matrix (L) was obtained

using QL = ς 2 BB T as the process noise matrix where ς = 6 is the LTR gain, and

RL = diag[I2 , 0.01] as the measurement noise matrix. The resulting gain matrices

122
 10 From δe(t) to q(t)
From δe(t) to Az(t)

Magnitude [dB]
0

−10

−20

−30

−40
−1 0 1 2
10 10 10 10
Frequency [rad/sec]

50
Phase [deg]

−50

−100

−150

−1 0 1 2
10 10 10 10
Frequency [rad/sec]

Figure 69: Frequency response of the linearized model

for this design are:

 
K = 18.55 3.09 −0.24 12.45 −0.02 0.02 −3.53 (231)

 T
 0.45 12.48 1.45 0.21 5.42 −2.25 −0.23 
 
L = 
 −0.81 −16.29 1.73 0.29 −42.10 10.77 0.53 
 (232)
 
−3.58 −23.13 −105.45 0.01 −168.11 31.89 3.53

Figure 70 shows the frequency response of the loop transfer functions with the loop
broken at the plant input for both full-state feedback and LQG/LTR loops, where it

can be seen that the margins of the full-state feedback loop are nearly recovered.

We consider the command tracking problem with the initial conditions set zero.

Figures 71 and 72 show the responses with nominal control in the absence of uncer-

tainty, and without disturbances. The upper part of Figure 71 compares the pitch

rate command (qcmd (t)), the reference model response (qm (t)), and the actual plant

123
 30
20
Full−State Feedback
Output Feedback

Magnitude [dB]
10
0
−10
−20
−30
−40
−50

−1 0 1 2
10 10 10 10
Frequency [rad/sec]

−50

−100
Phase [deg]

−150

−200

−250

−300

−350
−1 0 1 2
10 10 10 10
Frequency [rad/sec]

Figure 70: Frequency response of the loop transfer functions with the loop broken
at the plant input for both full-state feedback and LQG/LTR loops

response (q(t)). The lower portion compares the command to the actuator (δe (t) cmd)

with the actuator response. The difference cannot be distinguished at this scale. The

upper left portion of Figure 72 shows that there is a significant amount of sensor

noise. The remainder of this figure shows the performance of the observer. The effect

of sensor noise is most evident in the rate of change of the torsional amplitude.

For the adaptive control design we used a combination of bias and sigmoidal basis
 
functions of the form β(x̂(t)) = 1, β1 (x̂1 (t)), . . . , β6 (x̂6 (t)) , where βi (x̂i (t)) =
1−e−x̂i (t)
√
1+e −x̂ i (t) , i = 1, . . . , 6. For this choice of basis functions, β̄ = 7. For a chosen

µ = 0.95, the upper bound for κ1 in (169) was computed as κ̄1 = 0.51. Furthermore,

for a chosen Q0 = 1.2I7 , it was determined by Lemma 2.1 that ν̄ = 228.7 which

results in the upper bound κ̄2 = 31.6 for κ2 in (224). Given these bounds, we set

Ω1 = 0.5I7 and κ2 = 5. Finally, we chose τ = 0.05 seconds.

124
 6
qcmd(t)

qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

5 δe(t)

−5

−10

−15

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 71: Pitch rate and elevator responses with nominal control in the absence of
uncertainty
Sensed Az (t) [g] & q(t) [deg/sec]

4
Az(t) sensed (x 10) α(t)
q(t) sensed
6 α(t) estim.
2 q(t) estim.
4
α(t) [deg]

0
2
−2
0
−4
−2
−6

−8 −4
0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

w(t) 3 dw(t)
]
√ ]

6 θ(t) dθ(t)
2s
w(t) [ f2t ] & θ(t) [ deg

deg
2

2
] & θ̇(t) [ √

w(t) estim. dw(t) estim.

4 θ(t) estim. 1 dθ(t) estim.
0
2
−1
0
ft
ẇ(t) [ 2s

−2

−2 −3

−4
−4
0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

Figure 72: Measurement and state responses and their estimates with nominal con-
trol in the absence of uncertainty

125
5.6.1 Nonlinear Uncertainty

We first consider the case when there exists a nonlinear uncertainty of the form

∆(x(t)) = 2.5α(t)q(t) + 2.5α2(t) + 1.5q 2 (t) + 1.25w(t)θ(t) + 0.25ẇ(t)θ̇(t)

+2.5θ(t)θ̇(t) − 2.5θ̇2 (t) + 0.00025ẇ 2(t) (233)

Figure 73 shows the responses with nominal control and Figure 74 shows the responses

with adaptive control. The response with nominal control eventually goes unstable.

The response with adaptation is stable, and tracking performance is nearly as good

as that observed in Figure 71 without uncertainty. Also, the level of noise present in

the elevator command for the adaptive result is comparable to that observed for the

nominal control without uncertainty in Figure 71.

6
qcmd(t)
qcmd (t) & q(t) [deg/s]

4
qm(t)
2 q(t)

−2

−4

−6

−8
0 2 4 6 8 10 12 14
Time [s]

10 δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

δe(t)
5

−5

−10

0 2 4 6 8 10 12 14
Time [s]

Figure 73: Pitch rate and elevator responses with nominal control for the case of
nonlinear uncertainty

126
 6
qcmd(t)
qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

10 δe(t)

−5

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 74: Pitch rate and elevator responses with adaptive control for the case of
nonlinear uncertainty

127
5.6.2 External Disturbance

Next we consider the case when a disturbance d(t) = 4sin(3t) is applied to the system

through the control input. Figure 75 shows the responses with nominal control and
Figure 76 shows the responses with adaptive control. The tracking performance is

significantly improved with adaptation.

8
q (t)
cmd
qcmd (t) & q(t) [deg/s]

6
qm(t)
4
q(t)
2

−2

−4

−6

−8
0 2 4 6 8 10 12 14 16 18 20
Time [s]

δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

10
δe(t)
5

−5

−10

−15
0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 75: Pitch rate and elevator responses with nominal control for the case of
external disturbance

128
 6
qcmd(t)
qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

10 δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

δe(t)
5

−5

−10

−15

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 76: Pitch rate and elevator responses with adaptive control for the case of
external disturbance

129
5.6.3 Nonlinear Uncertainty and External Disturbance

The uncertainties in the previous two cases are combined in this example. Figure

77 shows that the response with adaptation remains stable, and adequate tracking
performance is still maintained.

6
qcmd(t)
qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

15
δe(t)
10

−5

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 77: Pitch rate and elevator responses with adaptive control for the case of
nonlinear uncertainty and external disturbance

130
5.6.4 Sudden Change in Ap

In this subsection, we illustrate an extreme case of parameter change, as might be

caused by structural damage, in which the second row of Ap changes sign at t = 6

seconds. Figure 78 shows that the response with nominal control becomes unstable,

whereas Figure 79 shows that the effect of this sudden change is hardly noticeable in

the response with adaptive control. We include Figure 80 to show that the estimation

performance of the observer is somewhat degraded by the uncertainty. In all the other

cases the effect of uncertainty on estimation performance was barely noticeable.

10 qcmd(t)
qcmd (t) & q(t) [deg/s]

qm(t)
5 q(t)

−5

−10

0 2 4 6 8 10 12 14 16 18 20
Time [s]

10 δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

δe(t)
5

−5

−10

−15

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 78: Pitch rate and elevator responses with nominal control for the case of
sudden change in the sign of second row of Ap

131
 6
qcmd(t)

qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

10
δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

δe(t)
5

−5

−10

−15

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 79: Pitch rate and elevator responses with adaptive control for the case of
sudden change in the sign of second row of Ap
Sensed Az (t) [g] & q(t) [deg/sec]

6
Az(t) sensed (x 10) α(t)
6
4
q(t) sensed α(t) estim.
2
q(t) estim. 4
α(t) [deg]

0 2

−2 0

−4
−2
−6
−4
−8
0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

4
w(t) dw(t)
]
√ ]

6 θ(t) 3 dθ(t)
2s
w(t) [ f2t ] & θ(t) [ deg

deg
2

] & θ̇(t) [ √

w(t) estim. 2 dw(t) estim.

4
θ(t) estim. dθ(t) estim.
1
2 0

0 −1
ft
ẇ(t) [ 2s

−2
−2
−3
−4 −4

0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

Figure 80: Measurement and state responses and their estimates with adaptive con-
trol for the case of sudden change in the sign of second row of Ap

132
5.6.5 Sudden Change in Cmα

Finally we illustrate additional two cases. For the first case the pitch stiffness, Cmα ,

becomes zero at t = 6 seconds. Figure 81 shows that the response with nominal
control, whereas Figure 82 shows that the effect of this sudden change is hardly

noticeable in the response with adaptive control. Figure 83 shows that the estimation

performance of the observer is also as good as that observed in Figure 72 without

uncertainty. For the second case Cmα changes sign at t = 6 seconds. Figure 84

shows that the response with nominal control becomes unstable, whereas Figure 85

shows that the tracking performance is nearly as good as that observed in Figure 71
without uncertainty. Once again, Figure 86 shows that the estimation performance

of the observer is as good as that observed in Figure 72.

6
qcmd(t)
qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

15
δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

10 δe(t)
5

−5

−10

−15

−20

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 81: Pitch rate and elevator responses with nominal control for the case of
sudden change in Cmα

133
 6
q (t)
cmd

qcmd (t) & q(t) [deg/s]

4 qm(t)
q(t)
2

−2

−4

−6
0 2 4 6 8 10 12 14 16 18 20
Time [s]

δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

10
δe(t)
5

−5

−10

−15

−20
0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 82: Pitch rate and elevator responses with adaptive control for the case of
sudden change in Cmα
Sensed Az (t) [g] & q(t) [deg/sec]

4
Az(t) sensed (x 10) α(t)
6
q(t) sensed α(t) estim.
2 q(t) estim.
4
α(t) [deg]

0
2
−2
0
−4
−2
−6
−4
−8
0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

4
w(t) dw(t)
]
√ ]

6 θ(t) dθ(t)
2s
w(t) [ f2t ] & θ(t) [ deg

deg
2

] & θ̇(t) [ √

w(t) estim. 2 dw(t) estim.

4 θ(t) estim. dθ(t) estim.
0
2

0 −2
ft
ẇ(t) [ 2s

−2
−4
−4

0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

Figure 83: Measurement and state responses and their estimates with adaptive con-
trol for the case of sudden change in Cmα

134
 qcmd(t)

qcmd (t) & q(t) [deg/s]

25
qm(t)
20
q(t)
15

10

−5
0 2 4 6 8 10 12
Time [s]

20
δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

0
δe(t)
−20

−40

−60

−80

−100

0 2 4 6 8 10 12
Time [s]

Figure 84: Pitch rate and elevator responses with nominal control for the case of
sudden change in the sign of Cmα

5
qcmd(t)
qcmd (t) & q(t) [deg/s]

qm(t)
q(t)

−5
0 2 4 6 8 10 12 14 16 18 20
Time [s]

15 δe(t) cmd.
δe (t) cmd. & δe (t) [deg]

10 δe(t)
5

−5

−10

−15

−20

0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 85: Pitch rate and elevator responses with adaptive control for the case of
sudden change in the sign of Cmα

135
 Sensed Az (t) [g] & q(t) [deg/sec]

A (t) sensed (x 10) α(t)

4 z
6
q(t) sensed α(t) estim.
2 q(t) estim.
4

α(t) [deg]
0
2
−2
0
−4
−2
−6
−4
−8
0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

4
w(t) dw(t)
]
√ ]

6 θ(t) dθ(t)
2s
w(t) [ f2t ] & θ(t) [ deg

deg
2

2
] & θ̇(t) [ √

w(t) estim. dw(t) estim.

4 θ(t) estim. dθ(t) estim.
0
2

0 −2
ft
ẇ(t) [ 2s

−2
−4

−4
−6
0 5 10 15 20 0 5 10 15 20
Time [s] Time [s]

Figure 86: Measurement and state responses and their estimates with adaptive con-
trol for the case of sudden change in the sign of Cmα

136
5.7 Conclusion

In this chapter we extended derivative-free adaptive control to the output feedback

case. The level of complexity of the output feedback form is far less than many other

methods, and it can be implemented in a form that augments an observer based linear

controller architecture. Illustrative examples on a model of wing rock dynamics and

an aeroelastic model of longitudinal dynamics of a generic transport model confirm

that this is a promising approach, particularly for applications in which there is a

combination of modeling uncertainty and external disturbances affecting the plant,

or there is uncertainty in the aeroelastic model of the aircraft.

137
 CHAPTER VI

Concluding Remarks and Future Research

6.1 Concluding Remarks

The intent of this dissertation has been to present new approaches to improve standard

designs in adaptive control theory and novel adaptive control architectures. For this

purpose, we first present a procedure that approximately enforces linear constraints

to modify an existing adaptive control design by using a Kalman filter optimization

approach in Chapter II. One key difference in the Kalman filter based approach to

modification is that the resulting gain on the modification term is optimal and time
varying. The proposed approach to deriving a modification term for an adaptive con-

troller can be used in place of all modification terms that can be equivalently viewed

as the gradient of a norm measure of a linear constraint on the adaptation gains. We

then present a Kalman filter based adaptive controller approach in Chapter III. This

approach is an extension of the results in Chapter II where instead of optimizing the

modification gain, the adaptation gain is optimized. This approach results in a time-
varying adaptation gain. The simulation results on a model of wing rock dynamics

illustrate the presented theory and show significant improvement over other gradient

based modification approaches.

The second major contribution of this dissertation is the development of a novel

derivative-free, delayed weight update law for adaptive control in Chapter IV. The

key feature is that the stability analysis is performed under the assumption that the
ideal weights are bounded, but otherwise arbitrarily time varying. The approach is

particularly useful for situations in which the nature of the system uncertainty cannot

be adequately represented by a set of basis functions with unknown constant weights.

138
The system error signals, including the state tracking error and the weight update
error, are proven to be uniformly ultimately bounded using a Lyapunov-Krasovskii

functional without the introduction of modification terms. We have shown that the

errors approach the ultimate bound exponentially in time. We have also provided

an analysis that shows how the design parameters in the adaptive law can influence

the ultimate bound, and the exponential rate of convergence to the bound. We

have also shown how the derivative-free adaptive law can be modified using many
of the existing modification terms. These results have been extended to the case of

uncertainty in control effectiveness. We show various examples that illustrate the

superior performance of derivative free adaptation in comparison to a conventional

adaptive law, when it is evaluated for a variety of cases in which there is a sudden

or rapidly varying set of dynamics. Furthermore, the performance of the proposed

approach is successfully evaluated in flight tests performed by NASA on AirSTAR

and the results are analyzed in Section IV.8.

Finally, we develop a novel approach for extending all the methods developed

in this dissertation to the case of output feedback in Chapter V. The approach is

developed only for the case of derivative-free adaptive control, and the extension

of the other approaches developed previously for the state feedback case to output

feedback is left as a future research topic. The level of complexity of the output
feedback form is far less than many other methods, and it can be implemented in

a form that augments an observer based linear controller architecture. Illustrative

examples on a model of wing rock dynamics and an aeroelastic model of longitudinal

dynamics of a generic transport model confirm that this is a promising approach.

6.2 Recommended Future Research

We recommend the following future research topics: (i) The results of Chapters II

and III can be extended to output feedback adaptive control. This can be done by

139
using the results presented in Chapter V. (ii) The results of Chapters II and III can
be used to develop a Kalman filter based derivative-free adaptive control law. This

will result in an optimal and time-varying adaptation gain and allow to achieve a

given performance criteria without the need for excessive adaptation gain tuning.

(iii) In Chapter IV, we showed that the closed loop system errors are uniformly

ultimately bounded and approach to the ultimate bound exponentially in time. This

characterizes the transient and steady-state performance bounds for the controlled
uncertain system by derivative-free adaptive control law. Next step towards this

direction is to characterize the performance bounds for the output feedback case

in Chapter V. (iv) We suggest an extension of the derivative-free output feedback

adaptive control law to the case when there is an input uncertainty in the system.

This extension can be done similar to the extension given in Chapter IV.5. (v) An

extension for the results of Chapters IV and V is to develop a derivative-free adaptive

control law for decentralized control. (vi) The results in this dissertation assume that

the full dimension of the plant is known. Therefore, they do not apply to the problem

of unmodeled dynamics. This problem is an important extension for all the results

in this dissertation.

Quantification of transient dynamics and gain/time-delay sensitivity at the sys-

tem inputs and outputs are two open problems in adaptive control theory. To address
these problems, there is a need to develop a systematic procedure to assign rates of

adaptation and the design matrices used in the Lyapunov or Riccati equations. As

discussed, derivative-free adaptive control is shown to have guaranteed transient and

steady-state performance bounds. However, there is still a need to quantify the tran-

sient dynamics. Towards this end, we suggest the following future research topic:

(vii) An analysis based on Singular Perturbation Theory is used to quantify the tran-
sient dynamics of an adaptively controlled system in Ref. 52. A similar analysis can

be applied to derivative-free adaptive control which should provide a similar result.

140
In order to address gain/time-delay sensitivity, we combined derivative-free adaptive
control with adaptive loop recovery modification as proposed in Ref. 10, where this

modification term results in approximate retention of reference model loop properties

such as gain and time-delay margins. However, we did this combination only for

the full-state feedback control problem. Therefore, we suggest the following future

research topic: (viii) Adaptive loop recovery modification term can be employed with

the proposed derivative-free output feedback adaptive controller in order to improve

gain and time-delay margins of the adaptively controlled system.

141
 REFERENCES

[1] “Generic transport model documentation for release v0903,” in NASA Langley
Research Center, 2009.
[2] Ananthakrishnan, S., “Adaptive tachometer feedback augmentation of the
shuttle remote manipulator control system,” 1995.
[3] Anderson, C., Hittle, D., Katz, A., and Kretchmar, R., “Synthesisof
reinforcement learning, neural networks and pi control applied to a simulated
heating coil,” Artificial Intelligence in Engineering, vol. 11, pp. 421–429, 1997.
[4] Astrom, K. J. and Wittenmark, B., Adaptive Control. Englewood Cliffs,
NJ: Prentice-Hall, 1994.
[5] Bernstein, D. S., Matrix Mathematics: Theory, Facts, and Formulas. Prince-
ton, NJ: Princeton University Press, 2nd ed., 2009.
[6] Bodson, M. and Groszkiewicz, J., “Multivariable adaptive algorithms for re-
configurable flight control,” IEEE Transactions on Control Systems Technology,
vol. 5, no. 2, pp. 217–229, 1997.
[7] Bordignon, K. and Durham, W., “Closed form solutions to the constrained
control allocation problem,” in AIAA Guidance, Navigation, and Control Con-
ference, (Scottsdale, AZ), 1993.
[8] Boyd, S. and Vandenberghe, L., Convex optimization. Cambridge, NY:
Cambridge University Press, 2004.
[9] Butchart, R. L. and Shackcloth, B., “Synthesis of model reference adap-
tive control systems by lyapunov’s second method,” Proc. IFAC Symp. on Adap-
tive Control, Teddington, UK, 1965.
[10] Calise, A. and Yucelen, T., “Adaptive loop recovery,” AIAA Journal of
Guidance, Control, and Dynamics, (submitted).
[11] Calise, A. J., Hovakimyan, N., and Idan, M., “Adaptive output feedback
control of nonlinear systems using neural networks,” Automatica, vol. 37, no. 8,
pp. 1201–1211, 2001.
[12] Calise, A. J. and Rysdyk, R., “Nonlinear adaptive flight control using neural
networks,” IEEE Control Systems Magazine, vol. 18, 1998.
[13] Calise, A. J., Shin, Y., and Johnson, M. D., “Comparison study of classical
and neural network based adaptive control of wing rock,” Proc. AIAA Guid.,
Navig., and Contr. Conf., 2004.

142
[14] Chandramohan, R., Yucelen, T., Calise, A. J., Chowdhary, G., and
Johnson, E. N., “Experimental results for kalman filter modification in adap-
tive control,” Proc. AIAA Infotech Conf., 2010.

[15] Chandramohan, R., Yucelen, T., Calise, A. J., Chowdhary, G., and
Johnson, E. N., “Flight test results for kalman filter and H2 modification in
adaptive control,” Proc. AIAA Guid., Navig., and Contr. Conf., 2010.

[16] Chandramohan, R., Yucelen, T., Calise, A. J., and Johnson, E. N.,
“Experimental evaluation of derivative-free model reference adaptive control,” in
AIAA Guidance Navigation and Control Conference, 2010.

[17] Chen, W. and Chowdhury, F. N., “Simultaneous identification of time-

varying parameters and estimation of system states using iterative learning ob-
servers,” Int. J. Sys. Science, vol. 38, pp. 39–45, 2007.

[18] Choi, J. and Farrell, J., “Observer-based backstepping control using on-line
approximation,” in Proceedings of the American Control Conference, pp. 3646–
3650, 2000.

[19] Choukroun, D., Weiss, H., Bar-Itzhack, I. Y., and Oshman, Y., Kalman
filtering for matrix estimation. University of California Postprints, 2006.

[20] Chowdhary, G. and Johnson, E. N., “Theory and flight-test validation of

a concurrent-learning adaptive controller,” AIAA Journal of Guidance, Control,
and Dynamics, 2011.

[21] Dash, P., Elangovan, S., and Liew, A., “Design of nonlinear expert su-
pervisory controllers for power system stabilization,” Eleictric Power Systems
Research, vol. 33, 1995.

[22] Doyle, J., Glover, K., Khargonekar, P., and Francis, B., “State-space
solutions to standard H2 and H∞ problems,” IEEE Trans. Autom. Control,
vol. 34, pp. 831–847, 1989.

[23] Doyle, J. and Stein, G., “Robustness with observers,” IEEE Trans. Autom.
Control, vol. 24, pp. 607–611, 1979.

[24] Durham, W., “Constrained control allocation,” in AIAA Journal of Guidance,

Control, and Dynamics, vol. 16, pp. 717–725, 1993.

[25] Egardt, B., Stability of Adaptive Controllers. Berlin: Springer-Verlag, 1979.

[26] Egardt, B., “Unification of some continuous-time adaptive control schemes,”

IEEE Trans. Autom. Control, vol. 24, pp. 588–592, 1979.

[27] Esfandiari, F. and Khalil, H. K., “Output-feedback stabilization of fully

linearizable systems,” International Journal of Control, vol. 56, no. 5, pp. 1007–
1037, 1992.

143
[28] Goodwin, G. C. and Mayne, D. Q., “A parameter estimation perspective of
continuous time model reference adaptive control,” Automatica, vol. 23, pp. 57–
70, 1987.
[29] Goodwin, G. C. and Sin, K. S., Adaptive Filtering, Prediction, and Control.
Englewood Cliffs, NJ: Prentice-Hall, 1984.
[30] Haddad, W. M. and Chellaboina, V., Nonlinear Dynamical Systems and
Control. A Lyapunov-Based Approach. Princeton, NJ: Princeton University
Press, 2008.
[31] Hale, J. and Lunel, S. M. V., Introduction to Functional Differential Equa-
tions. New York, NY: Springer-Verlag, 1993.
[32] Hoagg, J. B., Santillo, M. A., and Bernstein, D. S., “Discrete-time
adaptive command following and disturbance rejection with unknown exogenous
dynamics,” IEEE Trans. Autom. Control, vol. 53, pp. 912–928, 2008.
[33] Hovakimyan, N. and Cao, C., L1 Adaptive Control: Guaranteed Robustness
with Fast Adaptation. Philadelphia, PA: Society for Industrial and Applied Math-
ematics, 2010.
[34] Hovakimyan, N., Nardi, F., Calise, A., and Kim, N., “Adaptive output
feedback control of uncertain nonlinear systems using single-hidden-layer neural
networks,” IEEE Transactions on Neural Networks, vol. 13, no. 6, pp. 1420–1431,
2002.
[35] Idan, M., Johnson, M., and Calise, A. J., “Hierarchical approach to adap-
tive control for improved flight safety,” AIAA Journal of Guidance, Control, and
Dynamics, vol. 25, 2002.
[36] Ioannou, P. and Fidan, B., Adaptive Control Tutorial. Philadelphia, PA:
Society for Industrial and Applied Mathematics, 2006.
[37] Ioannou, P. and Kokotovic, P., “Instability analysis and improvement of
robustness of adaptive control,” Automatica, vol. 20, no. 5, pp. 583–594, 1984.
[38] Ioannou, P. A. and Sun, J., Robust Adaptive Control. Englewood Cliffs, NJ:
Prentice-Hall, 1995.
[39] Johnson, E. N., Limited authority adaptive flight control. School of Aerospace
Engineering, Georgia Institute of Technology, 2000.
[40] Johnson, E. N. and Calise, A. J., Limited authority adaptive flight control
for reusable launch vehicles. Journal of Guidance, Control, and Dynamics, vol.
26, pp. 906–913, 2003.
[41] Jordan, T. L. and Bailey, R. M., “NASA Langley’s airstar testbed: A sub-
scale flight test capability for flight dynamics and control system experiments,”
in Proc. AIAA Guidance, Navigation, and Control Conf., 2008.

144
[42] Jordan, T. L., Foster, J. V., Bailey, R. M., and Belcastro, C. M.,
“Airstar: A uav platform for flight dynamics and control system testing,” in
Proc. AIAA Aerodyn. Meas. Tech. and Ground Testing Conf., 2006.

[43] Kalman, R. E., “Contributions to the theory of optimal control,” Bol. Soc.
Matem. Mex., pp. 102–119, 1960.

[44] Kaufman, H., Barkana, I., and Sobel, K., Direct Adaptive Control Algo-
rithms: Theory and Applications. London: Springer-Verlag, 1993.

[45] Khalil, H. K., Nonlinear Systems, 2nd ed. Upper Saddle River, NJ: Prentice
Hall, 1993.

[46] Kim, K., Yucelen, T., and Calise, A., “A parameter dependent riccati
equation approach to output feedback adaptive control,” in AIAA Guidance,
Navigation, and Control Conference, 2011 (submitted).

[47] Kim, K., Yucelen, T., and Calise, A. J., “K modification in adaptive con-
trol,” Proc. AIAA Infotech Conf., 2010.

[48] Kim, Y. and Lewis, F., High Level Feedback Control with Neural Networks.
NJ: World Scientific, 1998.

[49] Krstic, M., Kanellakopoulos, I., and Kokotovic, P., Nonlinear and
Adaptive Control Design. New York: Wiley, 1995.

[50] Lavretsky, E., “Combined / composite model reference adaptive control,”

Proc. AIAA Guid., Navig., and Contr. Conf., 2009.

[51] Lavretsky, E., “Adaptive output feedback design using asymptotic properties
of LQG/LTR controllers,” in AIAA Guidance, Navigation, and Control Confer-
ence, 2010.

[52] Lavretsky, E., “Reference dynamics modification in adaptive controllers for

improved transient performance,” in AIAA Guidance, Navigation, and Control
Conference, 2011.

[53] Li, D., Hovakimyan, N., and Cao, C., “L1 adaptive control in the presence
of input saturation,” Proc. AIAA Guid., Navig., and Contr. Conf., 2009.

[54] Lyapunov, A. M., The General Problem of the Stability of Motion. Kharkov,
Russia: Kharkov Mathematical Society, 1892.

[55] Lyapunov, A. M., General Problem on Stability of Motion. Moscow:

Grostechizdat, 1935.

[56] Marino, R. and Tomei, P., Nonlinear Control Design: Geometric, Adaptive
and Robust. Englewood Cliffs, NJ: Prentice-Hall, 1995.

145
[57] Monopoli, R. V., “Model reference adaptive control with an augmented error
signal,” IEEE Trans. Autom. Control, vol. 19, pp. 474–484, 1974.

[58] Morse, A. S., “Global stability of parameter adaptive control schemes,” IEEE
Trans. Autom. Control, vol. 25, pp. 433–439, 1980.

[59] N. Nguyen, I. Tuzcu, T. Y. and Calise, A. J., “Longitudinal dynamics

and adaptive control application for an aeroelastic generic transport model,” in
AIAA Guidance, Navigation and Control Conference, 2011.

[60] Narendra, K. S. and Annaswamy, A. M., “A new adaptive law for robust
adaptation without persistent excitation,” IEEE Trans. Autom. Control, vol. 32,
no. 2, pp. 134–145, 1987.

[61] Narendra, K. S. and Annaswamy, A. M., Stable Adaptive Systems. Engle-

wood Cliffs, NJ: Prentice-Hall, 1989.

[62] Narendra, K. S., Annaswamy, A. M., and Singh, R. P., “A general ap-
proach to the stability of adaptive systems,” Int. J. Control, vol. 41, pp. 193–216,
1985.

[63] Nguyen, N., Krishnakmar, K., and Boskovic, J., “An optimal control
modification to model-reference adaptive control for fast adaptation,” Proc.
AIAA Guid., Navig., and Contr. Conf., 2008.

[64] Osburn, P. V., Whitaker, H. P., and Kezer, A., “New developments in the
design of adaptive control systems,” Paper No 61-39, Institute of Aeronautical
Sciences, 1961.

[65] Parks, P. C., “Lyapunov redesign of model reference adaptive control systems,”
IEEE Trans. Autom. Control, vol. 11, pp. 362–365, 1966.

[66] Pomet, J. B. and Praly, L., “Adaptive nonlinear regulation: Estimation

from lyapunov equation,” IEEE Transactions on Automatic Control, vol. 37,
pp. 729–740, 1992.

[67] Potter, J. E., “Matrix quadratic solutions,” SIAM Journal on Applied Math-
ematics, vol. 14, pp. 496–501, 1966.

[68] Sastry, S. and Bodson, M., Adaptive Control: Stability, Convergence, and
Robustness. Englewood Cliffs, NJ: Prentice-Hall, 1989.

[69] Seshagiri, S. and Khalil, H. K., “Output feedback control of nonlinear sys-
tems using rbf neural networks,” IEEE Transactions on Neural Networks, vol. 11,
no. 1, pp. 69–79, 2000.

[70] Singh, S. N., Yim, W., and Wells, W. R., “Direct adaptive and neural con-
trol of wing-rock motion of slender delta wings,” Journal of Guidance, Control,
and Dynamics, vol. 18, pp. 25–30, 1995.

146
[71] Slotine, J. J. E. and Li, W., Applied Nonlinear Control. Englewood Cliffs,
NJ: Prentice-Hall, 1991.

[72] Spooner, J., Maggiore, M., Ordonez, R., and Passino, K., Stable Adap-
tive Control and Estimation for Nonlinear Systems: Neural and Fuzzy Approxi-
mator Techniques. New York, NY: John Wiley and Sons, 2002.

[73] Tao, G., Adaptive Control Design and Analysis. New York, NY: Wiley, 2003.

[74] Volyanskyy, K., Haddad, W., and Calise, A., “A new neuroadaptive con-
trol arichitecture for nonlinear uncertain dynamical systems: Beyond sigma -
and e - modifications,” IEEE Transactions on Neural Networks, vol. 20, no. 11,
pp. 1707–1723, 2009.

[75] Volyanskyy, K. Y., Calise, A. J., and Yang, B. J., “A novel Q modifica-
tion term for adaptive control,” Proc. IEEE Amer. Contr. Conf., 2006.

[76] Volyanskyy, K. Y., Calise, A. J., Yang, B. J., and Lavretsky, E., “An
error minimization method in adaptive control,” Proc. AIAA Guid., Navig., and
Contr. Conf., 2006.

[77] Volyanskyy, K. Y., Haddad, W. M., and Calise, A. J., “A new neuroad-
aptive control architecture for nonlinear uncertain dynamical systems: Beyond
σ- and e- modifications,” IEEE Trans. on Neural Networks, 2010.

[78] Wise, K. A., “Applied controls research topics in the aerospace industry,” in
IEEE Conference on Decision and Control, (New Orleans, LA), 1995.

[79] Wise, K. A., “A trade study on missile autopilot design using optimal control
theory,” in AIAA Guidance, Navigation and Control Conference, 2007.

[80] Wohletz, J., Paduano, J., and Annaswamy, A., “Retrofit systems for re-
configuration in civil aviation,” in Proceedings of the AIAA Guidance, Naviga-
tion, and Control Conference, (Portland, OR), 1999.

[81] Yucelen, T. and Calise, A. J., “Enforcing a linear constraint in adaptive

control: A kalman filter optimization approach,” Proc. AIAA Guid., Navig.,
and Contr. Conf., 2009.

[82] Yucelen, T. and Calise, A. J., “Kalman filter modification in adaptive con-
trol,” AIAA J. of Guid., Contr., and Dyn., vol. 33, pp. 426–439, 2010.

[83] Yucelen, T. and Calise, A. J., “Kalman filter modification in adaptive con-
trol,” AIAA Journal of Guidance, Control, and Dynamics, vol. 33, pp. 426–439,
2010.

[84] Yucelen, T. and Calise, A. J., “Adaptive control with loop transfer recovery:
A kalman filter approach,” Proc. IEEE Amer. Contr. Conf., 2011 (to appear).

147
[85] Yucelen, T. and Calise, A. J., “Derivative-free model reference adaptive
control,” in AIAA Journal of Guidance, Control, and Dynamics, 2011 (to ap-
pear).

[86] Yucelen, T. and Calise, A. J., “Experimental evaluation of derivative-free

adaptive controller with loop transfer recovery on the nasa airstar flight test
vehicle,” in AIAA Infotech Conf., (St. Louis, MO), 2011 (to appear).

[87] Yucelen, T. and Calise, A. J., “Derivative-free model reference adaptive

control,” in AIAA Guid., Nav., and Contr. Conf., (Toronto, ON), August 2010.

[88] Yucelen, T. and Calise, A. J., “Derivative-free model reference adaptive

control of a generic transport model,” in AIAA Guid., Nav., and Contr. Conf.,
(Toronto, ON), August 2010.

[89] Yucelen, T. and Calise, A. J., “Derivative-free model reference adaptive

control,” AIAA J. Guid., Contr., Dyn., (to appear).

[90] Yucelen, T., Calise, A. J., Haddad, W. M., and Volyanskyy, K. Y.,
“A comparison of a new neuroadaptive controller architecture with the σ and e
modification architectures,” Proc. AIAA Guid., Navig., and Contr. Conf., 2009.

[91] Yucelen, T., Haddad, W. M., and Calise, A. J., “A neuroadaptive control
architecture for nonlinear uncertain dynamical systems with input constraints,”
IEEE Trans. on Neural Networks, 2011 (to appear).

148
 VITA

Tansel Yucelen received the B.S. degree in control engineering from Istanbul Technical
University, Istanbul, Turkey, in 2006, and the M.S. degree in electrical and computer

engineering from Southern Illinois University, Carbondale, Illinois, in 2008. He is

currently working towards the Ph.D. degree in aerospace engineering at the School of

Aerospace Engineering, Georgia Institute of Technology, Atlanta, Georgia.

His research interests include active noise control, active vibration control, adap-

tive control, adaptive systems, automation, autonomous systems, computational meth-

ods, control applications, delay systems, direct adaptive control, discrete-time sys-

tems, flexible structures, flight control, fuzzy control, H2 control, H∞ control, identi-

fication, indirect adaptive control, linear matrix inequalities, linear systems, manned

aerial vehicles, mechanical systems, modeling, neural networks, neuroadaptive con-

trol, nonlinear systems, optimal control, optimization, real-time systems, robotics,

robust adaptive control, robust control, stability of linear systems, stability of non-
linear systems, time-varying systems, uncertain systems, unmanned aerial vehicles.

He is a student member of the American Institute of Aeronautics and Astronautics

(AIAA) and the Institute of Electrical and Electronics Engineers (IEEE).

149